[async_tp] Base support ag-transpose-mm(mat_B) case

ghstack-source-id: edd51b9c46e46e8eca0c45e0ea53c1b26b375c01
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163069

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -5,9 +5,9 @@ GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
 # Set CUDA architecture lists to match x86 build_cuda.sh
 if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
+    export TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;8.0;9.0"
 elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+    export TORCH_CUDA_ARCH_LIST="7.0;8.0;9.0;10.0;12.0"
 elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi
@@ -31,7 +31,8 @@ pip install -r /pytorch/requirements.txt
 pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
@@ -41,9 +42,13 @@ else
         echo "Bundling CUDA libraries with wheel for aarch64."
     else
         echo "Using nvidia libs from pypi for aarch64."
+        # Fix platform constraints in PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64
+        # Replace 'platform_machine == "x86_64"' with 'platform_machine == "aarch64"'
+        export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${PYTORCH_EXTRA_INSTALL_REQUIREMENTS//platform_machine == \'x86_64\'/platform_machine == \'aarch64\'}"
         echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
         export USE_NVIDIA_PYPI_LIBS=1
     fi
 
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -138,8 +138,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     folder = os.path.dirname(wheel_path)
     os.mkdir(f"{folder}/tmp")
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Delete original wheel since it will be repackaged
-    os.system(f"rm {wheel_path}")
 
     # Check if we should use PyPI NVIDIA libraries or bundle system libraries
     use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
@@ -213,8 +211,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         ]
 
         # CUDA version-specific libraries
-        if "13" in desired_cuda:
-            minor_version = desired_cuda[-1]
+        if "130" in desired_cuda:
             version_specific_libs = [
                 "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
                 "/usr/local/cuda/lib64/libcublas.so.13",
@@ -224,7 +221,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/usr/local/cuda/lib64/libcusolver.so.12",
                 "/usr/local/cuda/lib64/libnvJitLink.so.13",
                 "/usr/local/cuda/lib64/libnvrtc.so.13",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
+                "/usr/local/cuda/lib64/libnvrtc-builtins.so.13.0",
             ]
         elif "12" in desired_cuda:
             # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
@@ -240,8 +237,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/usr/local/cuda/lib64/libnvrtc.so.12",
                 f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
             ]
-        else:
-            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
 
         # Combine all libraries
         libs_to_copy = common_libs + version_specific_libs
@@ -280,7 +275,14 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+        repaired_wheel_name = wheel_name.replace(
+            "linux_aarch64", "manylinux_2_28_aarch64"
+        )
+        print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
+        os.rename(
+            f"/{folder}/dist/{wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -317,7 +319,7 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = ""
+    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "

.ci/docker/build.sh

@@ -214,7 +214,8 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
-    ANACONDA_PYTHON_VERSION=3.10
+    # TODO (huydhn): Upgrade this to Python >= 3.10
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes

.ci/docker/centos-rocm/Dockerfile

@@ -56,13 +56,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+56392aa978594cc155fa8af48cd949f5b5f1823a

.ci/docker/ci_commit_pins/rocm-composable-kernel.txt

@@ -1 +0,0 @@
-7fe50dc3da2069d6645d9deb8c017a876472a977

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-5ae38bdb0dc066c5823e34dc9797afb9de42c866
+fccfc522864cf8bc172abe0cd58ae5581e2d44b9

.ci/docker/common/install_executorch.sh

@@ -42,27 +42,22 @@ install_pip_dependencies() {
   # A workaround, ExecuTorch has moved to numpy 2.0 which is not compatible with the current
   # numba and scipy version used in PyTorch CI
   conda_run pip uninstall -y numba scipy
-  # Yaspin is needed for running CI test (get_benchmark_analysis_data.py)
-  pip_install yaspin==3.1.0
 
   popd
 }
 
 setup_executorch() {
+  pushd executorch
+
   export PYTHON_EXECUTABLE=python
-  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON -DEXECUTORCH_BUILD_TESTS=ON"
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
 
   as_jenkins .ci/scripts/setup-linux.sh --build-tool cmake || true
+  popd
 }
 
-if [ $# -eq 0 ]; then
-  clone_executorch
-  install_buck2
-  install_conda_dependencies
-  install_pip_dependencies
-  pushd executorch
-  setup_executorch
-  popd
-else
-  "$@"
-fi
+clone_executorch
+install_buck2
+install_conda_dependencies
+install_pip_dependencies
+setup_executorch

.ci/docker/common/install_rocm.sh

@@ -2,11 +2,6 @@
 
 set -ex
 
-# for pip_install function
-source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
-
-ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
-
 ver() {
     printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
 }
@@ -118,8 +113,6 @@ EOF
         rm -rf HIP clr
     fi
 
-    pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
     # Cleanup
     apt-get autoclean && apt-get clean
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@@ -183,8 +176,6 @@ install_centos() {
       sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
   done
 
-  pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
   # Cleanup
   yum clean all
   rm -rf /var/cache/yum

.ci/docker/ubuntu-rocm/Dockerfile

@@ -52,13 +52,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/libtorch/build.sh

@@ -7,4 +7,4 @@ set -ex
 
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.10" ${SCRIPTPATH}/../manywheel/build.sh
+USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh

.ci/lumen_cli/cli/lib/core/vllm/vllm_build.py

@@ -66,11 +66,6 @@ class VllmBuildParameters:
         "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
     )
 
-    # the cleaning script to remove torch dependencies from pip
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
     output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
 
@@ -165,7 +160,6 @@ class VllmBuildRunner(BaseRunner):
         logger.info("Running vllm build with inputs: %s", inputs)
         vllm_commit = clone_vllm()
 
-        self.cp_torch_cleaning_script(inputs)
         self.cp_dockerfile_if_exist(inputs)
         # cp torch wheels from root direct to vllm workspace if exist
         self.cp_torch_whls_if_exist(inputs)
@@ -211,11 +205,6 @@ class VllmBuildRunner(BaseRunner):
         copy(inputs.torch_whls_path, tmp_dir)
         return tmp_dir
 
-    def cp_torch_cleaning_script(self, inputs: VllmBuildParameters):
-        script = get_path(inputs.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
     def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
         if not inputs.use_local_dockerfile:
             logger.info("using vllm default dockerfile.torch_nightly for build")

.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py

@@ -11,7 +11,7 @@ from typing import Any
 
 from cli.lib.common.cli_helper import BaseRunner
 from cli.lib.common.envs_helper import env_path_field, env_str_field, get_env
-from cli.lib.common.path_helper import copy, get_path, remove_dir
+from cli.lib.common.path_helper import copy, remove_dir
 from cli.lib.common.pip_helper import (
     pip_install_first_match,
     pip_install_packages,
@@ -43,10 +43,6 @@ class VllmTestParameters:
 
     torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
 
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     def __post_init__(self):
         if not self.torch_whls_path.exists():
             raise ValueError("missing torch_whls_path")
@@ -96,13 +92,11 @@ class VllmTestRunner(BaseRunner):
         self._set_envs(params)
 
         clone_vllm(dst=self.work_directory)
-        self.cp_torch_cleaning_script(params)
         with working_directory(self.work_directory):
             remove_dir(Path("vllm"))
             self._install_wheels(params)
             self._install_dependencies()
         # verify the torches are not overridden by test dependencies
-
         check_versions()
 
     def run(self):
@@ -131,11 +125,6 @@ class VllmTestRunner(BaseRunner):
             # double check the torches are not overridden by other packages
             check_versions()
 
-    def cp_torch_cleaning_script(self, params: VllmTestParameters):
-        script = get_path(params.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
     def _install_wheels(self, params: VllmTestParameters):
         logger.info("Running vllm test with inputs: %s", params)
         if not pkg_exists("torch"):

.ci/pytorch/common_utils.sh

@@ -258,19 +258,11 @@ function install_torchrec_and_fbgemm() {
       git clone --recursive https://github.com/pytorch/fbgemm
       pushd fbgemm/fbgemm_gpu
       git checkout "${fbgemm_commit}" --recurse-submodules
-      # until the fbgemm_commit includes the tbb patch
-      patch <<'EOF'
---- a/FbgemmGpu.cmake
-+++ b/FbgemmGpu.cmake
-@@ -184,5 +184,6 @@ gpu_cpp_library(
-     fbgemm_gpu_tbe_cache
-     fbgemm_gpu_tbe_optimizers
-     fbgemm_gpu_tbe_utils
-+    tbb
-   DESTINATION
-     fbgemm_gpu)
-EOF
-      python setup.py bdist_wheel --build-variant=rocm
+      python setup.py bdist_wheel \
+        --build-variant=rocm \
+        -DHIP_ROOT_DIR="${ROCM_PATH}" \
+        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
       popd
 
       # Save the wheel before cleaning up

.ci/pytorch/functorch_doc_push_script.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# This is where the local pytorch install in the docker image is located
+pt_checkout="/var/lib/jenkins/workspace"
+source "$pt_checkout/.ci/pytorch/common_utils.sh"
+echo "functorch_doc_push_script.sh: Invoked with $*"
+
+set -ex -o pipefail
+
+version=${DOCS_VERSION:-nightly}
+echo "version: $version"
+
+# Build functorch docs
+pushd $pt_checkout/functorch/docs
+make html
+popd
+
+git clone https://github.com/pytorch/functorch -b gh-pages --depth 1 functorch_ghpages
+pushd functorch_ghpages
+
+if [ "$version" == "main" ]; then
+  version=nightly
+fi
+
+git rm -rf "$version" || true
+mv "$pt_checkout/functorch/docs/build/html" "$version"
+
+git add "$version" || true
+git status
+git config user.email "soumith+bot@pytorch.org"
+git config user.name "pytorchbot"
+# If there aren't changes, don't make a commit; push is no-op
+git commit -m "Generate Python docs from pytorch/pytorch@${GITHUB_SHA}" || true
+git status
+
+if [[ "${WITH_PUSH:-}" == true ]]; then
+  git push -u origin gh-pages
+fi
+
+popd

.ci/pytorch/numba-cuda-13.patch

@@ -1,25 +0,0 @@
-From 6e08c9d08e9de59c7af28b720289debbbd384764 Mon Sep 17 00:00:00 2001
-From: Michael Wang <13521008+isVoid@users.noreply.github.com>
-Date: Tue, 1 Apr 2025 17:28:05 -0700
-Subject: [PATCH] Avoid bumping certain driver API to avoid future breakage
- (#185)
-
-Co-authored-by: isVoid <isVoid@users.noreply.github.com>
----
- numba_cuda/numba/cuda/cudadrv/driver.py | 3 +++
- 1 file changed, 3 insertions(+)
-
-diff --git a/numba_cuda/numba/cuda/cudadrv/driver.py b/numba_cuda/numba/cuda/cudadrv/driver.py
-index 1641bf77..233e9ed7 100644
---- a/numba_cuda/numba/cuda/cudadrv/driver.py
-+++ b/numba_cuda/numba/cuda/cudadrv/driver.py
-@@ -365,6 +365,9 @@ def _find_api(self, fname):
-         else:
-             variants = ('_v2', '')
- 
-+        if fname in ("cuCtxGetDevice", "cuCtxSynchronize"):
-+            return getattr(self.lib, fname)
-+
-         for variant in variants:
-             try:
-                 return getattr(self.lib, f'{fname}{variant}')

.ci/pytorch/smoke_test/smoke_test.py

@@ -386,8 +386,8 @@ def smoke_test_compile(device: str = "cpu") -> None:
 
 
 def smoke_test_nvshmem() -> None:
-    if not torch.cuda.is_available() or target_os == "windows":
-        print("Windows platform or CUDA is not available, skipping NVSHMEM test")
+    if not torch.cuda.is_available():
+        print("CUDA is not available, skipping NVSHMEM test")
         return
 
     # Check if NVSHMEM is compiled in current build
@@ -396,9 +396,7 @@ def smoke_test_nvshmem() -> None:
     except ImportError:
         # Not built with NVSHMEM support.
         # torch is not compiled with NVSHMEM prior to 2.9
-        from torch.torch_version import TorchVersion
-
-        if TorchVersion(torch.__version__) < (2, 9):
+        if torch.__version__ < "2.9":
             return
         else:
             # After 2.9: NVSHMEM is expected to be compiled in current build

.ci/pytorch/test.sh

@@ -32,16 +32,6 @@ if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /v
   git config --global --add safe.directory /var/lib/jenkins/workspace
 fi
 
-
-# Patch numba to avoid CUDA-13 crash, see https://github.com/pytorch/pytorch/issues/162878
-NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
-if [ -n "$NUMBA_CUDA_DIR" ]; then
-  NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
-  pushd "$NUMBA_CUDA_DIR"
-  patch -p4 <"$NUMBA_PATCH"
-  popd
-fi
-
 echo "Environment variables:"
 env
 
@@ -1550,10 +1540,14 @@ test_executorch() {
   install_torchvision
   install_torchaudio
 
-  INSTALL_SCRIPT="$(pwd)/.ci/docker/common/install_executorch.sh"
-
   pushd /executorch
-  "${INSTALL_SCRIPT}" setup_executorch
+
+  export PYTHON_EXECUTABLE=python
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
+
+  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
+  # from the PR
+  bash .ci/scripts/setup-linux.sh --build-tool cmake
 
   echo "Run ExecuTorch unit tests"
   pytest -v -n auto
@@ -1567,6 +1561,10 @@ test_executorch() {
 
   popd
 
+  # Test torchgen generated code for Executorch.
+  echo "Testing ExecuTorch op registration"
+  "$BUILD_BIN_DIR"/test_edge_op_registration
+
   assert_git_not_dirty
 }
 
@@ -1574,7 +1572,6 @@ test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
         test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
-        distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1724,6 +1721,11 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
+  if [[ "${SHARD_NUMBER}" == 1 ]]; then
+    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
+      test_inductor_distributed
+    fi
+  fi
 elif [[ "${TEST_CONFIG}" == *einops* ]]; then
   test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then

.ci/pytorch/windows/internal/driver_update.bat

@@ -1,9 +1,9 @@
-set WIN_DRIVER_VN=580.88
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
-curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
+set WIN_DRIVER_VN=528.89
+set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
+curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
 if errorlevel 1 exit /b 1
 
-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1
 
-del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL

.ci/wheel/build_wheel.sh

@@ -85,7 +85,7 @@ mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
 # Create an isolated directory to store this builds pytorch checkout and conda
 # installation
 if [[ -z "$MAC_PACKAGE_WORK_DIR" ]]; then
-    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_${DESIRED_PYTHON}_$(date +%H%M%S)"
+    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_conda_${DESIRED_PYTHON}_$(date +%H%M%S)"
 fi
 mkdir -p "$MAC_PACKAGE_WORK_DIR" || true
 if [[ -n ${GITHUB_ACTIONS} ]]; then
@@ -96,11 +96,11 @@ fi
 whl_tmp_dir="${MAC_PACKAGE_WORK_DIR}/dist"
 mkdir -p "$whl_tmp_dir"
 
-mac_version='macosx-11_0-arm64'
+mac_version='macosx_11_0_arm64'
 libtorch_arch='arm64'
 
 # Create a consistent wheel package name to rename the wheel to
-wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version//[-,]/_}.whl"
+wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version}.whl"
 
 ###########################################################
 
@@ -125,6 +125,7 @@ popd
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
 export MACOSX_DEPLOYMENT_TARGET=11.0
+export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
@@ -132,19 +133,25 @@ RENAME_WHEEL=true
 case $desired_python in
     3.14t)
         echo "Using 3.14 deps"
-        mac_version='macosx-11.0-arm64'
         NUMPY_PINNED_VERSION="==2.1.0"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
         RENAME_WHEEL=false
         ;;
     3.14)
         echo "Using 3.14t deps"
-        mac_version='macosx-11.0-arm64'
         NUMPY_PINNED_VERSION="==2.1.0"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
         RENAME_WHEEL=false
         ;;
     3.13t)
         echo "Using 3.13 deps"
         NUMPY_PINNED_VERSION="==2.1.0"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+        desired_python="3.13"
         RENAME_WHEEL=false
         ;;
     3.13)
@@ -169,12 +176,17 @@ case $desired_python in
         ;;
 esac
 
+# Install into a fresh env
+tmp_env_name="wheel_py$python_nodot"
+conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
+source activate "$tmp_env_name"
+
 PINNED_PACKAGES=(
     "numpy${NUMPY_PINNED_VERSION}"
 )
-python -mvenv ~/${desired_python}-build
-source ~/${desired_python}-build/bin/activate
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
+retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
+pip install requests ninja typing-extensions
+retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
 # For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
@@ -188,7 +200,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-_PYTHON_HOST_PLATFORM=${mac_version} ARCHFLAGS="-arch arm64" python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name "${mac_version//[-.]/_}"
+python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name ${mac_version}
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.flake8

@@ -73,7 +73,7 @@ exclude =
     ./docs/src,
     ./functorch/docs,
     ./functorch/examples,
-    ./functorch/docs/source/tutorials,
+    ./functorch/notebooks,
     ./scripts,
     ./test/generated_type_hints_smoketest.py,
     ./third_party,

.github/actionlint.yaml

@@ -21,7 +21,6 @@ self-hosted-runner:
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
-    - linux.arm64.r7g.12xlarge.memory
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu

.github/actions/reuse-old-whl/reuse_old_whl.py

@@ -264,7 +264,7 @@ def unzip_artifact_and_replace_files() -> None:
         change_content_to_new_version(f"artifacts/dist/{old_stem}/torch/version.py")
 
         for file in Path(f"artifacts/dist/{old_stem}").glob(
-            "*.dist-info/*",
+            "*.dist-info/**",
         ):
             change_content_to_new_version(file)
 

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+27fc2493d383354a008106f22f3be232badee9a1

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +1 @@
-08ae0af1395c8d8471f4025deb6af9aef90b342f
+7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-5963b98b465007e3cfb0d39447e4459a8afa96dc
+e10fef08838612b4560e9c72e5cb1414a5edfa13

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-c77852e117bdf056c8e9a087e51d6f65cf6ba53d
+6c5478ff7c3d50dd1e3047d72ec5909bea474073

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -82,10 +82,16 @@ RUN if command -v apt-get >/dev/null; then \
         apt-get update -y \
         && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git curl wget sudo vim; \
     fi \
     && python3 --version && python3 -m pip --version
 
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version >/dev/null 2>&1; then \
@@ -214,16 +220,11 @@ ARG SCCACHE_S3_NO_CREDENTIALS=0
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..."; \
-        if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-aarch64-unknown-linux-musl"; \
-        else \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-x86_64-unknown-linux-musl"; \
-        fi; \
-        curl -L -o sccache.tar.gz "https://github.com/mozilla/sccache/releases/download/v0.8.1/${SCCACHE_ARCHIVE}.tar.gz" \
+        echo "Installing sccache..." \
+        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
         && tar -xzf sccache.tar.gz \
-        && sudo mv "${SCCACHE_ARCHIVE}"/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz "${SCCACHE_ARCHIVE}" \
+        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
+        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
         && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
         && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
@@ -284,7 +285,7 @@ RUN if command -v apt-get >/dev/null; then \
         && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
         && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git curl wget sudo vim; \
     fi \
     && python3 --version && python3 -m pip --version
 
@@ -297,6 +298,12 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version > /dev/null 2>&1; then \

.github/ci_configs/vllm/use_existing_torch.py

@@ -1,17 +0,0 @@
-import glob
-
-
-requires_files = glob.glob("requirements/*.txt")
-requires_files += ["pyproject.toml"]
-for file in requires_files:
-    print(f">>> cleaning {file}")
-    with open(file) as f:
-        lines = f.readlines()
-    if "torch" in "".join(lines).lower():
-        print("removed:")
-        with open(file, "w") as f:
-            for line in lines:
-                if "torch" not in line.lower():
-                    f.write(line)
-    print(f"<<< done cleaning {file}")
-    print()

.github/requirements/pip-requirements-macOS.txt

@@ -15,7 +15,7 @@ optree==0.13.0
 packaging==23.1
 parameterized==0.8.1
 pillow==10.3.0
-protobuf==5.29.5
+protobuf==5.29.4
 psutil==5.9.8
 pygments==2.15.0
 pytest-cpp==2.3.0
@@ -26,7 +26,7 @@ pytest-xdist==3.3.1
 pytest==7.3.2
 pyyaml==6.0.2
 scipy==1.12.0
-setuptools==78.1.1
+setuptools==72.1.0
 sympy==1.13.3
 tlparse==0.4.0
 tensorboard==2.13.0

.github/scripts/docathon-label-sync.py

@@ -39,9 +39,7 @@ def main() -> None:
     pull_request_label_names = [label.name for label in pull_request_labels]
     issue_label_names = [label.name for label in issue_labels]
     labels_to_add = [
-        label
-        for label in issue_label_names
-        if label not in pull_request_label_names and label != "actionable"
+        label for label in issue_label_names if label not in pull_request_label_names
     ]
     if not labels_to_add:
         print("The pull request already has the same labels.")

.github/scripts/generate_binary_build_matrix.py

@@ -43,55 +43,55 @@ CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
-        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.2.1 | "

.github/scripts/prepare_vllm_wheels.sh

@@ -1,94 +0,0 @@
-#!/usr/bin/env bash
-
-set -eux
-
-torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-nightly=$(echo ${torch_version} | cut -d'.' -f4)
-
-# Copied from .ci/manywheel/build_common.sh
-make_wheel_record() {
-  fpath=$1
-  if echo $fpath | grep RECORD >/dev/null 2>&1; then
-    echo "$fpath,,"
-  else
-    fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
-    fsize=$(ls -nl $fpath | awk '{print $5}')
-    echo "$fpath,sha256=$fhash,$fsize"
-  fi
-}
-
-change_wheel_version() {
-  local package=$1
-  local wheel=$2
-  local f_version=$3
-  local t_version=$4
-
-  # Extract the wheel
-  ${PYTHON_EXECUTABLE} -mwheel unpack $wheel
-
-  mv "${package}-${f_version}" "${package}-${t_version}"
-  # Change the version from f_version to t_version in the dist-info dir
-  pushd "${package}-${t_version}"
-  mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
-
-  pushd "${package}-${t_version}.dist-info"
-  sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
-
-  # Update the version in METADATA and its SHA256 hash
-  sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
-  # then add PyTorch nightly dependency of vLLM
-  if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
-    sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
-  fi
-  sed -i '/METADATA,sha256/d' RECORD
-  popd
-
-  make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
-  popd
-
-  # Repack the wheel
-  ${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
-
-  # Clean up
-  rm -rf "${package}-${t_version}"
-}
-
-repackage_wheel() {
-  local package=$1
-  pushd $package
-
-  local orig_wheel=$(find . -name *${package//-/_}*)
-  local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-
-  local version=""
-  if [[ "${package}" == vllm ]]; then
-    # Copied from vllm/.buildkite/scripts/upload-wheels.sh
-    version=1.0.0
-  else
-    version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
-  fi
-  local nightly_version=$version.$nightly
-
-  # Use nightly version
-  change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
-  # Clean up
-  rm "${orig_wheel}"
-
-  auditwheel repair --plat $PLATFORM *.whl \
-    --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
-  local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
-  local repair_wheel=$(basename ${repair_wheel})
-  popd
-
-  cp ${package}/wheelhouse/${repair_wheel} .
-  rm -rf $package
-}
-
-# Require to re-package the wheel
-${PYTHON_EXECUTABLE} -mpip install wheel==0.45.1
-
-pushd externals/vllm/wheels
-for package in xformers flashinfer-python vllm; do
-  repackage_wheel $package
-done
-popd

.github/templates/macos_binary_build_workflow.yml.j2

@@ -22,16 +22,6 @@ name: !{{ build_environment }}
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
 {%- endmacro %}
 
-{%- macro setup_python(py_ver) -%}
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
-          freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
-{%- endmacro %}
-
 on:
 # TODO: Migrate to new ciflow trigger, reference https://github.com/pytorch/pytorch/pull/70321
   push:
@@ -71,13 +61,23 @@ jobs:
     {%- endif %}
     steps:
       !{{ set_runner_specific_vars() }}
-      !{{ setup_python(config.get("python_version", "3.10")) }}
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       !{{ common.checkout(deep_clone=False, directory="pytorch") }}
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -94,6 +94,8 @@ jobs:
 {%- if config["package_type"] == "wheel" %}
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -104,9 +106,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086

.github/workflows/_docs.yml

@@ -75,6 +75,10 @@ jobs:
             runner: ${{ inputs.runner_prefix }}linux.2xlarge
             # It takes less than 30m to finish python docs unless there are issues
             timeout-minutes: 30
+          - docs_type: functorch
+            runner: ${{ inputs.runner_prefix }}linux.2xlarge
+            # It takes less than 15m to finish functorch docs unless there are issues
+            timeout-minutes: 15
     # Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)
     # The current name requires updating the database last docs push query from test-infra every time the matrix is updated
     name: build-docs-${{ matrix.docs_type }}-${{ inputs.push }}
@@ -207,6 +211,16 @@ jobs:
           path: cppdocs/
           s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/cppdocs
 
+      - name: Upload functorch Docs Preview
+        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
+        if: ${{ github.event_name == 'pull_request' && matrix.docs_type == 'functorch' && steps.build-docs.outcome == 'success' }}
+        with:
+          retention-days: 14
+          s3-bucket: doc-previews
+          if-no-files-found: error
+          path: functorch_ghpages/nightly/
+          s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/functorchdocs
+
       - name: Teardown Linux
         uses: pytorch/test-infra/.github/actions/teardown-linux@main
         if: always()

.github/workflows/_linux-test.yml

@@ -169,7 +169,7 @@ jobs:
         id: install-nvidia-driver
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
         with:
-          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '580.82.07' }}
+          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '570.133.07' }}
         if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
 
       - name: Setup GPU_FLAG for docker run

.github/workflows/_rocm-test.yml

@@ -62,11 +62,6 @@ on:
         required: false
         type: number
         default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 env:
   GIT_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
 
@@ -81,9 +76,10 @@ jobs:
     strategy:
       matrix: ${{ fromJSON(inputs.test-matrix) }}
       fail-fast: false
-    runs-on: ${{ matrix.runner }}
     timeout-minutes: ${{ matrix.mem_leak_check == 'mem_leak_check' && 600 || inputs.timeout-minutes }}
+    runs-on: ${{ matrix.runner }}
     steps:
+      # [see note: pytorch repo ref]
       - name: Checkout PyTorch
         uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
         with:
@@ -135,9 +131,6 @@ jobs:
 
       - name: Start monitoring script
         id: monitor-script
-        if: ${{ !inputs.disable-monitor }}
-        shell: bash
-        continue-on-error: true
         env:
           JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
@@ -145,6 +138,9 @@ jobs:
           WORKFLOW_RUN_ID: ${{github.run_id}}
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
+        if: ${{ !inputs.disable-monitor }}
+        shell: bash
+        continue-on-error: true
         run: |
           python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
@@ -182,12 +178,6 @@ jobs:
         run: |
           echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
 
-      - name: Preserve github env variables for use in docker
-        shell: bash
-        run: |
-          env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-          env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-
       - name: Test
         id: test
         env:
@@ -203,22 +193,20 @@ jobs:
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
           BRANCH: ${{ steps.parse-ref.outputs.branch }}
           SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
-          TEST_CONFIG: ${{ matrix.config }}
-          SHARD_NUMBER: ${{ matrix.shard }}
-          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
-          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
           TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}
           NO_TEST_TIMEOUT: ${{ steps.keep-going.outputs.ci-no-test-timeout }}
           NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
+          TEST_CONFIG: ${{ matrix.config }}
+          SHARD_NUMBER: ${{ matrix.shard }}
+          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
+          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
           DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           set -x
@@ -248,7 +236,6 @@ jobs:
             -e GITHUB_RUN_ATTEMPT \
             -e JOB_ID \
             -e JOB_NAME \
-            -e BASE_SHA \
             -e BRANCH \
             -e SHA1 \
             -e AWS_DEFAULT_REGION \
@@ -266,12 +253,10 @@ jobs:
             -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
-            -e HUGGING_FACE_HUB_TOKEN \
             -e DASHBOARD_TAG \
             --env-file="${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \
-            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --security-opt seccomp=unconfined \
             --cap-add=SYS_PTRACE \
             --shm-size="8g" \

.github/workflows/build-vllm-wheel.yml

@@ -12,9 +12,6 @@ on:
     paths:
       - .github/workflows/build-vllm-wheel.yml
       - .github/ci_commit_pins/vllm.txt
-  schedule:
-    # every morning at 01:30PM UTC, 9:30AM EST, 6:30AM PST
-    - cron: 30 13 * * *
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
@@ -27,33 +24,21 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
+        # TODO (huydhn): Add cu130 https://github.com/pytorch/pytorch/pull/162000#issuecomment-3261541554
         device: [ 'cu128', 'cu129' ]
+        runner: [ 'linux.12xlarge.memory' ]
         include:
-          - platform: manylinux_2_28_x86_64
-            device: cu128
+          - device: cu128
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu129
+          - device: cu129
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu128
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
-            runner: linux.arm64.r7g.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu129
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
-            runner: linux.arm64.r7g.12xlarge.memory
-    name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
+    name: "Build ${{ matrix.device }} vLLM wheel"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
     env:
       PY_VERS: ${{ matrix.python-version }}
       MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
-      PLATFORM: ${{ matrix.platform }}
+      PLATFORM: 'manylinux_2_28_x86_64'
       BUILD_DEVICE: ${{ matrix.device }}
     steps:
       - name: Setup SSH (Click me for login details)
@@ -74,6 +59,20 @@ jobs:
         run: |
           set -eux
 
+          # Keep PyTorch nightly wheel here so that we can install it later during
+          # vLLM build process
+          mkdir -p "${RUNNER_TEMP}/artifacts/"
+
+          container_name=$(docker run \
+            --tty \
+            --detach \
+            -e PLATFORM \
+            -v "${GITHUB_WORKSPACE}:/pytorch" \
+            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
+            -w /artifacts/ \
+            "${MANYLINUX_IMAGE}"
+          )
+
           # Determine python executable for given version (copied from build-triton-wheel)
           case $PY_VERS in
           3.10)
@@ -103,21 +102,6 @@ jobs:
             ;;
           esac
 
-          # Keep PyTorch nightly wheel here so that we can install it later during
-          # vLLM build process
-          mkdir -p "${RUNNER_TEMP}/artifacts/"
-
-          container_name=$(docker run \
-            --tty \
-            --detach \
-            -e PLATFORM \
-            -e PYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" \
-            -v "${GITHUB_WORKSPACE}:/pytorch" \
-            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
-            -w /artifacts/ \
-            "${MANYLINUX_IMAGE}"
-          )
-
           docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip install \
             --pre torch torchvision torchaudio \
             --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
@@ -129,6 +113,7 @@ jobs:
             --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
 
           # Save this for later
+          echo "PYTHON_EXECUTABLE=${PYTHON_EXECUTABLE}" >> "$GITHUB_ENV"
           echo "container_name=${container_name}" >> "$GITHUB_ENV"
 
       - name: Build vLLM wheel
@@ -146,12 +131,41 @@ jobs:
           set -eux
 
           # Get these wheels ready, the vllm renaming logic is copied from its .buildkite/scripts/upload-wheels.sh
-          docker exec -t "${container_name}" bash -c /pytorch/.github/scripts/prepare_vllm_wheels.sh
+          docker exec -t "${container_name}" bash -c "
+            set -eux
+
+            nightly=\$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2 | cut -d'.' -f4)
+
+            pushd externals/vllm/wheels
+            for package in xformers flashinfer-python vllm; do
+              pushd \$package
+              auditwheel repair --plat \$PLATFORM *.whl \
+                --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
+              repair_wheel=\$(find wheelhouse -name *\${PLATFORM}*)
+              repair_wheel=\$(basename \${repair_wheel})
+              popd
+
+              cp \${package}/wheelhouse/\${repair_wheel} .
+              version=\$(unzip -p \$repair_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+
+              if [[ \$package == vllm ]]; then
+                new_wheel=\${repair_wheel/\$version/1.0.0.\$nightly}
+              else
+                major_version=\$(echo \$version | tr '.+' '.' | cut -d'.' -f1-3)
+                new_wheel=\${repair_wheel/\$version/\$major_version.\$nightly}
+              fi
+
+              mv -- \$repair_wheel \$new_wheel
+              rm -rf \$package
+            done
+            popd
+          "
+
           docker exec -t "${container_name}" chown -R 1000:1000 /artifacts
 
       - uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
         with:
-          name: vllm-wheel-${{ matrix.device }}-${{ matrix.platform }}-${{ matrix.python-version }}
+          name: vllm-wheel-${{ matrix.device }}-${{ matrix.python-version }}-${{ env.PLATFORM }}
           if-no-files-found: error
           path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
 
@@ -161,29 +175,27 @@ jobs:
 
   # Copied from build-triton-wheel workflow (mostly)
   upload-wheel:
-    name: "Upload ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
+    name: "Upload ${{ matrix.device }} vLLM wheel"
     needs:
       - build-wheel
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
       matrix:
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
         device: [ 'cu128', 'cu129' ]
     env:
-      PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}
     permissions:
       id-token: write
       contents: read
     container:
       image: continuumio/miniconda3:4.12.0
-    environment: ${{ ((github.event_name == 'push' && github.event.ref == 'refs/heads/main') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && 'nightly-wheel-upload' || '' }}
+    environment: ${{ (github.event_name == 'push' && github.event.ref == 'refs/heads/main') && 'nightly-wheel-upload' || '' }}
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Configure AWS credentials(PyTorch account) for main
-        if: ${{ (github.event_name == 'push' && github.event.ref == 'refs/heads/main') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
+        if: ${{ github.event_name == 'push' && github.event.ref == 'refs/heads/main' }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_nightly_build_wheels
@@ -207,15 +219,15 @@ jobs:
         run: |
           set -eux
           mkdir -p "${RUNNER_TEMP}/artifacts/"
-          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-"${PLATFORM}"-*/* "${RUNNER_TEMP}/artifacts/"
+          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-*/* "${RUNNER_TEMP}/artifacts/"
 
-      - name: Set DRY_RUN
-        if: ${{ (github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v'))) || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
+      - name: Set DRY_RUN (only for tagged pushes)
+        if: ${{ github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v')) }}
         shell: bash
         run: |
           echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
 
-      - name: Set UPLOAD_CHANNEL
+      - name: Set UPLOAD_CHANNEL (only for tagged pushes)
         if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
         shell: bash
         run: |

.github/workflows/docker-builds.yml

@@ -71,7 +71,8 @@ jobs:
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
-          pytorch-linux-jammy-py3-clang12-executorch,
+          # Executorch pin needs update
+          # pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu,
           pytorch-linux-noble-riscv64-py3.12-gcc14
         ]

.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml

@@ -132,7 +132,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -178,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -335,7 +335,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +381,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -427,7 +427,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -538,7 +538,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -584,7 +584,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -741,7 +741,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -787,7 +787,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -833,7 +833,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -944,7 +944,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -990,7 +990,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1036,7 +1036,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1147,7 +1147,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1193,7 +1193,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1239,7 +1239,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1350,7 +1350,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1396,7 +1396,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1442,7 +1442,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/generated-linux-binary-manywheel-main.yml

@@ -60,7 +60,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing

.github/workflows/generated-linux-binary-manywheel-nightly.yml

@@ -127,7 +127,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_6-test:  # Testing
@@ -193,7 +193,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-test:  # Testing
@@ -259,7 +259,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -719,7 +719,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_6-test:  # Testing
@@ -785,7 +785,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-test:  # Testing
@@ -851,7 +851,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -1311,7 +1311,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_6-test:  # Testing
@@ -1377,7 +1377,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing
@@ -1443,7 +1443,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -1903,7 +1903,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_6-test:  # Testing
@@ -1969,7 +1969,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-test:  # Testing
@@ -2035,7 +2035,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2495,7 +2495,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_6-test:  # Testing
@@ -2561,7 +2561,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-test:  # Testing
@@ -2627,7 +2627,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -3087,7 +3087,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_6-test:  # Testing
@@ -3153,7 +3153,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_8-test:  # Testing
@@ -3219,7 +3219,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -3679,7 +3679,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_6
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_6-test:  # Testing
@@ -3745,7 +3745,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_8-test:  # Testing
@@ -3811,7 +3811,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda13_0-test:  # Testing

.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml

@@ -60,13 +60,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.10.4"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -81,9 +81,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"

.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml

@@ -56,13 +56,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.10.4"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -77,9 +77,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -95,6 +99,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -105,9 +111,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -166,13 +196,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.11.4"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -187,9 +217,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -205,6 +239,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -215,9 +251,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -276,13 +336,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.12.4"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -297,9 +357,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -315,6 +379,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -325,9 +391,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -386,13 +476,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.13.4"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -407,9 +497,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -425,6 +519,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -435,9 +531,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -496,13 +616,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.13.4"
-          freethreaded: true
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -517,9 +637,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -535,6 +659,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -545,9 +671,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -606,13 +756,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
-          freethreaded: false
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -627,9 +777,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -645,6 +799,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -655,9 +811,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086
@@ -716,13 +896,13 @@ jobs:
           echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
           # shellcheck disable=SC2129
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
-          freethreaded: true
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
       - name: Checkout PyTorch
         uses: actions/checkout@v4
         with:
@@ -737,9 +917,13 @@ jobs:
         working-directory: pytorch
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -755,6 +939,8 @@ jobs:
           "${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -765,9 +951,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086

.github/workflows/inductor-nightly.yml

@@ -37,7 +37,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-default-label-prefix
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       test-matrix: |
@@ -56,7 +56,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: nightly-dynamo-benchmarks-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.nightly-dynamo-benchmarks-build.outputs.docker-image }}
       test-matrix: ${{ needs.nightly-dynamo-benchmarks-build.outputs.test-matrix }}
       timeout-minutes: 720

.github/workflows/inductor-perf-test-nightly-x86-zen.yml

@@ -43,11 +43,6 @@ on:
         required: false
         type: boolean
         default: false
-      freezing:
-        description: Run freezing?
-        required: false
-        type: boolean
-        default: true
       benchmark_configs:
         description: The list of configs used the benchmark
         required: false
@@ -80,7 +75,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -106,8 +101,8 @@ jobs:
     needs: inductor-build
     if: github.event.schedule == '0 7 * * *'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
+      build-environment: linux-jammy-py3.9-gcc11-build
+      dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
       timeout-minutes: 720
@@ -121,9 +116,10 @@ jobs:
     name: inductor-test
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-build
+    if: github.event_name == 'workflow_dispatch'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      dashboard-tag: training-${{ inputs.training || 'false' }}-inference-${{ inputs.inference || 'true' }}-default-${{ inputs.default || 'true' }}-dynamic-${{ inputs.dynamic || 'true' }}-cppwrapper-${{ inputs.cppwrapper || 'true' }}-aotinductor-${{ inputs.aotinductor || 'true' }}-freezing-${{ inputs.freezing || 'true' }}
+      build-environment: linux-jammy-py3.9-gcc11-build
+      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
       timeout-minutes: 720

.github/workflows/inductor-perf-test-nightly-x86.yml

@@ -80,7 +80,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -107,7 +107,7 @@ jobs:
     needs: inductor-build
     if: github.event.schedule == '0 7 * * *'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
@@ -124,7 +124,7 @@ jobs:
     needs: inductor-build
     if: github.event_name == 'workflow_dispatch'
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-freezing-${{ inputs.freezing }}
       docker-image: ${{ needs.inductor-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}

.github/workflows/inductor-periodic.yml

@@ -39,7 +39,7 @@ jobs:
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
-      cuda-arch-list: '8.0;8.6'
+      cuda-arch-list: '8.6'
       test-matrix: |
         { include: [
           { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
@@ -62,7 +62,7 @@ jobs:
           { config: "dynamic_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "dynamic_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "dynamic_inductor_torchbench", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
+          { config: "aot_inductor_huggingface", shard: 1, num_shards: 1, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_timm", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_timm", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
           { config: "aot_inductor_torchbench", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
@@ -154,7 +154,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-default-label-prefix
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       test-matrix: |
@@ -200,7 +200,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: periodic-dynamo-benchmarks-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor-unittest.yml

@@ -110,7 +110,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -127,7 +127,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/inductor.yml

@@ -79,7 +79,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
@@ -101,7 +101,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: inductor-cpu-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/nightly.yml

@@ -54,7 +54,7 @@ jobs:
       - get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image: ${{ needs.docs-build.outputs.docker-image }}
       push: ${{ github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' || startsWith(github.event.ref, 'refs/tags/v') }}
       run-doxygen: true

.github/workflows/operator_benchmark.yml

@@ -14,10 +14,6 @@ on:
   schedule:
     # Run at 07:00 UTC every Sunday
     - cron: 0 7 * * 0
-  pull_request:
-    paths:
-      - benchmarks/operator_benchmark/**
-      - .github/workflows/operator_benchmark.yml
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
@@ -33,7 +29,7 @@ jobs:
     name: opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -46,7 +42,7 @@ jobs:
     name: opbenchmark-on-demand-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -59,7 +55,7 @@ jobs:
     uses: ./.github/workflows/_linux-test.yml
     needs: opbenchmark-build
     with:
-      build-environment: linux-jammy-py3.10-gcc11-build
+      build-environment: linux-jammy-py3.9-gcc11-build
       docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/pull.yml

@@ -127,8 +127,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # More memory is needed to build with asan
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
@@ -318,6 +316,32 @@ jobs:
         ]}
     secrets: inherit
 
+  linux-jammy-py3-clang12-executorch-build:
+    if: false  # Docker build needs pin update
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-py3-clang12-executorch
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
+      test-matrix: |
+        { include: [
+          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-py3-clang12-executorch-test:
+    name: linux-jammy-py3-clang12-executorch
+    uses: ./.github/workflows/_linux-test.yml
+    needs: linux-jammy-py3-clang12-executorch-build
+    if: false # Has been broken for a while
+    with:
+      build-environment: linux-jammy-py3-clang12-executorch
+      docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
+    secrets: inherit
+
   linux-jammy-cuda12_8-py3_10-gcc9-inductor-build:
     name: cuda12.8-py3.10-gcc9-sm75
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/slow.yml

@@ -140,8 +140,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # More memory is needed to build with asan
-      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/trunk.yml

@@ -240,7 +240,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
@@ -255,31 +255,7 @@ jobs:
       - verify-cachebench-cpu-build
       - target-determination
     with:
-      build-environment: linux-jammy-py3.10-gcc11
+      build-environment: linux-jammy-py3.9-gcc11
       docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
       test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
     secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-build:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-executorch
-      test-matrix: |
-        { include: [
-          { config: "executorch", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-py3-clang12-executorch-test:
-    name: linux-jammy-py3-clang12-executorch
-    uses: ./.github/workflows/_linux-test.yml
-    needs: linux-jammy-py3-clang12-executorch-build
-    with:
-      build-environment: linux-jammy-py3-clang12-executorch
-      docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/vllm.yml

@@ -36,8 +36,6 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # When building vLLM, uv doesn't like that we rename wheel without changing the wheel metadata
-      allow-reuse-old-whl: false
       build-additional-packages: "vision audio"
       build-external-packages: "vllm"
       build-environment: linux-jammy-cuda12.8-py3.12-gcc11

.gitignore

@@ -259,9 +259,6 @@ gen
 .pytest_cache
 aten/build/*
 
-# Linker scripts for prioritized text optimization
-cmake/linker_script.ld
-
 # Bram
 plsdontbreak
 
@@ -392,5 +389,3 @@ android/pytorch_android_torchvision/.cxx
 
 # Claude Code local configuration
 CLAUDE.local.md
-/test_*.py
-/debug_*.py

.lintrunner.toml

@@ -13,7 +13,7 @@ exclude_patterns = [
     '**/fb/**',
     'functorch/docs/**',
     'functorch/examples/**',
-    'functorch/docs/source/tutorials/**',
+    'functorch/notebooks/**',
     'torch/_inductor/fx_passes/serialized_patterns/**',
     'torch/_inductor/autoheuristic/artifacts/**',
     'scripts/**',
@@ -964,6 +964,7 @@ exclude_patterns = [
     'test/jit/**',  # should be run through test/test_jit.py
     'test/ao/sparsity/**',  # should be run through test/test_ao_sparsity.py
     'test/fx/**',  # should be run through test/test_fx.py
+    'test/bottleneck_test/**',  # excluded by test/run_test.py
     'test/package/**',  # excluded by test/run_test.py
     'test/distributed/argparse_util_test.py',
     'test/distributed/bin/test_script.py',
@@ -1409,6 +1410,8 @@ exclude_patterns = [
     'torch/utils/benchmark/utils/timer.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/__init__.py',
     'torch/utils/benchmark/utils/valgrind_wrapper/timer_interface.py',
+    'torch/utils/bottleneck/__init__.py',
+    'torch/utils/bottleneck/__main__.py',
     'torch/utils/bundled_inputs.py',
     'torch/utils/checkpoint.py',
     'torch/utils/collect_env.py',
@@ -1565,6 +1568,7 @@ include_patterns = [
 exclude_patterns = [
     'caffe2/**',
     'functorch/docs/**',
+    'functorch/notebooks/**',
     'torch/_inductor/fx_passes/serialized_patterns/**',
     'torch/_inductor/autoheuristic/artifacts/**',
     'test/dynamo/cpython/**',

BUILD.bazel

@@ -810,7 +810,7 @@ cc_library(
     name = "torch_python",
     srcs = libtorch_python_core_sources
         + if_cuda(libtorch_python_cuda_sources)
-        + libtorch_python_distributed_sources
+        + if_cuda(libtorch_python_distributed_sources)
         + GENERATED_AUTOGRAD_PYTHON,
     hdrs = glob([
         "torch/csrc/generic/*.cpp",

CMakeLists.txt

@@ -234,7 +234,6 @@ cmake_dependent_option(INSTALL_TEST "Install test binaries if BUILD_TEST is on"
 option(USE_CPP_CODE_COVERAGE "Compile C/C++ with code coverage flags" OFF)
 option(USE_COLORIZE_OUTPUT "Colorize output during compilation" ON)
 option(USE_ASAN "Use Address+Undefined Sanitizers" OFF)
-option(USE_LSAN "Use Leak Sanitizer" OFF)
 option(USE_TSAN "Use Thread Sanitizer" OFF)
 option(USE_CUDA "Use CUDA" ON)
 option(USE_XPU "Use XPU" ON)
@@ -380,13 +379,6 @@ cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                        OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                         "CPU_AARCH64" OFF)
-# prioritized text linker, ON by default for AArch64+Linux, option visible to all AArch64, x86 and ppc64le.
-set(USE_PRIORITIZED_TEXT_DEFAULT OFF)
-if(LINUX AND CPU_AARCH64)
-  set(USE_PRIORITIZED_TEXT_DEFAULT ON)
-endif()
-cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker for ld."
-  "${USE_PRIORITIZED_TEXT_DEFAULT}" "CPU_INTEL OR CPU_AARCH64 OR CPU_POWER" OFF)
 
 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
@@ -664,11 +656,6 @@ endif(MSVC)
 
 string(APPEND CMAKE_CUDA_FLAGS " -Xfatbin -compress-all")
 
-# Set linker max-page-size to 64KiB on AArch64 Linux
-if(LINUX AND CPU_AARCH64)
-  add_link_options_if_supported("-z,max-page-size=0x10000")
-endif()
-
 # Set INTERN_BUILD_MOBILE for all mobile builds. Components that are not
 # applicable to mobile are disabled by this variable. Setting
 # `BUILD_PYTORCH_MOBILE_WITH_HOST_TOOLCHAIN` environment variable can force it
@@ -886,7 +873,7 @@ cmake_dependent_option(
   "Whether to build the flash_attention kernel for scaled dot product attention.\
   Will be disabled if not supported by the platform"
   ON
-  "(USE_CUDA AND NOT MSVC) OR USE_ROCM"
+  "USE_CUDA OR USE_ROCM;NOT MSVC"
   OFF)
 
 cmake_dependent_option(
@@ -902,9 +889,9 @@ IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
   set(USE_FBGEMM_GENAI off)
 endif()
 
-# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100.
-if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0" AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8 AND NOT WIN32)
-  message(STATUS "Setting USE_FBGEMM_GENAI to ON, doing CUDA build for SM100a")
+# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100
+if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0a")
+  message(WARNING "Setting USE_FBGEMM_GENAI to ON for CUDA build on SM100")
   set(USE_FBGEMM_GENAI ON)
 endif()
 
@@ -921,7 +908,7 @@ cmake_dependent_option(
 # USE_FLASH_ATTENTION -> USE_ROCM -> Dependencies.cmake -> aotriton.cmake
 #
 if(USE_ROCM)
-  if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
+  if(UNIX AND (USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION))
     include(cmake/External/aotriton.cmake)
   endif()
 endif()
@@ -1433,57 +1420,3 @@ if(BUILD_BUNDLE_PTXAS AND USE_CUDA)
   install(PROGRAMS "${PROJECT_BINARY_DIR}/ptxas"
           DESTINATION "${CMAKE_INSTALL_BINDIR}")
 endif()
-
-if(USE_PRIORITIZED_TEXT_FOR_LD)
-  add_compile_options(
-    $<$<COMPILE_LANGUAGE:C,CXX>:-ffunction-sections>
-    $<$<COMPILE_LANGUAGE:C,CXX>:-fdata-sections>
-  )
-  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
-  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
-
-  add_custom_command(
-    OUTPUT "${LINKER_SCRIPT_FILE_OUT}"
-    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py --filein "${LINKER_SCRIPT_FILE_IN}" --fout "${LINKER_SCRIPT_FILE_OUT}"
-    DEPENDS ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py "${LINKER_SCRIPT_FILE_IN}"
-    COMMENT "Generating prioritized text linker files"
-    VERBATIM
-  )
-
-  add_custom_target(generate_linker_script DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-
-  if(BUILD_PYTHON)
-    set(LINKER_OPT_TARGETS torch_python)
-  endif()
-
-  if(NOT BUILD_LIBTORCHLESS)
-    list(APPEND LINKER_OPT_TARGETS torch_cpu c10)
-    if(USE_CUDA)
-      list(APPEND LINKER_OPT_TARGETS torch_cuda c10_cuda)
-    endif()
-    if(USE_XPU)
-      list(APPEND LINKER_OPT_TARGETS torch_xpu c10_xpu)
-    endif()
-    if(USE_ROCM)
-      list(APPEND LINKER_OPT_TARGETS torch_hip c10_hip)
-    endif()
-  endif()
-
-  foreach(tgt IN LISTS LINKER_OPT_TARGETS)
-    if(TARGET ${tgt})
-      add_dependencies("${tgt}" generate_linker_script)
-      target_link_options_if_supported(${tgt} "-T,${LINKER_SCRIPT_FILE_OUT}")
-      set_property(TARGET ${tgt} APPEND PROPERTY LINK_DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
-    else()
-       message(WARNING "Requested target '${tgt}' for linker script optimization was not found.")
-    endif()
-  endforeach()
-
-else()
-  if(LINUX AND CPU_AARCH64)
-    message(WARNING [[
-    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
-    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
-    ]])
-  endif()
-endif()

RELEASE.md

@@ -50,7 +50,6 @@ Following is the Release Compatibility Matrix for PyTorch releases:
 
 | PyTorch version | Python | C++ | Stable CUDA | Experimental CUDA | Stable ROCm |
 | --- | --- | --- | --- | --- | --- |
-| 2.9 | >=3.10, <=(3.14, 3.14t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 13.0 (CUDNN 9.13.0.50) | ROCm 6.4 |
 | 2.8 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 12.9 (CUDNN 9.10.2.21) | ROCm 6.4 |
 | 2.7 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8 (CUDNN 9.1.0.70), CUDA 12.6 (CUDNN 9.5.1.17) | CUDA 12.8 (CUDNN 9.7.1.26) | ROCm 6.3 |
 | 2.6 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8, CUDA 12.4 (CUDNN 9.1.0.70) | CUDA 12.6 (CUDNN 9.5.1.17) | ROCm 6.2.4 |

SECURITY.md

@@ -16,8 +16,6 @@ However, if you believe you have found a security vulnerability in PyTorch, we e
 
 Please report security issues using https://github.com/pytorch/pytorch/security/advisories/new
 
-All reports submitted thru the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
-
 Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:
 
 https://www.facebook.com/whitehat

aten/src/ATen/CMakeLists.txt

@@ -265,14 +265,6 @@ IF(USE_FBGEMM_GENAI)
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
     list(FILTER fbgemm_genai_native_cuda_cu INCLUDE REGEX ${FBGEMM_CUTLASS_KERNELS_REGEX})
 
-    # PyTorch is not built for 10.0a in CI, due to lack of portability,
-    # so we need to explicitly build these files for 10.0a.
-    foreach(cu_file ${fbgemm_genai_native_cuda_cu})
-      _BUILD_FOR_ADDITIONAL_ARCHS(
-        "${cu_file}"
-        "100a")
-    endforeach()
-
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cpp
       "${FBGEMM_GENAI_SRCS}/common/*.cpp"
     )

aten/src/ATen/Context.cpp

@@ -180,7 +180,7 @@ void Context::setUserEnabledNNPACK(bool e) {
 }
 
 bool Context::allowTF32CuDNN(const std::string& op) const {
-  if (op.empty()){
+  if (op.size() == 0){
     bool allow_tf32_rnn = float32Precision("cuda", "rnn") == "tf32";
     bool allow_tf32_conv = float32Precision("cuda", "conv") == "tf32";
     TORCH_CHECK(
@@ -281,6 +281,9 @@ bool Context::userEnabledOverrideableSDP() const {
 
 static constexpr const auto cublas_config_var_name = "CUBLAS_WORKSPACE_CONFIG";
 static constexpr const std::array<const char*, 2> cublas_deterministic_configs = {":4096:8", ":16:8"};
+#ifdef USE_ROCM
+static constexpr const auto hipblaslt_allow_tf32 = "HIPBLASLT_ALLOW_TF32";
+#endif
 
 bool Context::checkCuBLASConfigDeterministic() {
   // If using CUDA 10.2 or greater, need to make sure CuBLAS workspace config
@@ -340,6 +343,12 @@ void Context::setImmediateMiopen(bool b) {
 }
 
 bool Context::allowTF32CuBLAS() const {
+#ifdef USE_ROCM
+    const auto allow_tf32 = c10::utils::check_env(hipblaslt_allow_tf32);
+    if (allow_tf32 != true) {
+      return false;
+    }
+#endif
   bool legacy_allow_tf32 = float32_matmul_precision != at::Float32MatmulPrecision::HIGHEST;
   bool allow_tf32_new = float32Precision("cuda", "matmul") == "tf32";
   TORCH_CHECK(
@@ -353,6 +362,14 @@ bool Context::allowTF32CuBLAS() const {
 }
 
 void Context::setAllowTF32CuBLAS(bool b) {
+#ifdef USE_ROCM
+  const auto allow_tf32 = c10::utils::check_env(hipblaslt_allow_tf32);
+  if (allow_tf32 != true) {
+    C10_LOG_FIRST_N(INFO, 10) << "torch.backends.cuda.matmul.allow_tf32 is not supported on ROCm by default. "
+                              << "Please set environment variable HIPBLASLT_ALLOW_TF32=1 to enable it.";
+    return;
+  }
+#endif
   float32_matmul_precision = b ? at::Float32MatmulPrecision::HIGH : at::Float32MatmulPrecision::HIGHEST;
   setFloat32Precision("cuda", "matmul", b ? "tf32" : "ieee");
 }
@@ -426,7 +443,7 @@ void Context::setFloat32Precision(const std::string& backend, const std::string&
     std::string msg;
     auto iterp = _fp32_precisions.find(backend);
     TORCH_CHECK(iterp != _fp32_precisions.end());
-    for (const auto& p : iterp->second) {
+    for (auto p : iterp->second) {
       msg += p;
       msg += " ";
     }

aten/src/ATen/DLConvertor.cpp

@@ -65,24 +65,14 @@ DLDataType getDLDataType(const Tensor& t) {
       break;
     // TODO(#146647): use macro here instead of spelling out each shell dtype
     case ScalarType::Float8_e5m2:
-      dtype.code = DLDataTypeCode::kDLFloat8_e5m2;
-      break;
     case ScalarType::Float8_e5m2fnuz:
-      dtype.code = DLDataTypeCode::kDLFloat8_e5m2fnuz;
-      break;
     case ScalarType::Float8_e4m3fn:
-      dtype.code = DLDataTypeCode::kDLFloat8_e4m3fn;
-      break;
     case ScalarType::Float8_e4m3fnuz:
-      dtype.code = DLDataTypeCode::kDLFloat8_e4m3fnuz;
-      break;
     case ScalarType::Float8_e8m0fnu:
-      dtype.code = DLDataTypeCode::kDLFloat8_e8m0fnu;
+      TORCH_CHECK_BUFFER(false, "float8 types are not supported by dlpack");
       break;
     case ScalarType::Float4_e2m1fn_x2:
-      dtype.code = DLDataTypeCode::kDLFloat4_e2m1fn;
-      dtype.lanes = 2;
-      dtype.bits = 4;
+      TORCH_CHECK_BUFFER(false, "float4 types are not supported by dlpack");
       break;
     case ScalarType::QInt8:
     case ScalarType::QUInt8:
@@ -187,11 +177,7 @@ static Device getATenDevice(DLDeviceType type, c10::DeviceIndex index, void* dat
 
 ScalarType toScalarType(const DLDataType& dtype) {
   ScalarType stype = ScalarType::Undefined;
-  if (dtype.code != DLDataTypeCode::kDLFloat4_e2m1fn) {
-    TORCH_CHECK_BUFFER(
-        dtype.lanes == 1,
-        "ATen does not support lanes != 1 for dtype code", std::to_string(dtype.code));
-  }
+  TORCH_CHECK_BUFFER(dtype.lanes == 1, "ATen does not support lanes != 1");
   switch (dtype.code) {
     case DLDataTypeCode::kDLUInt:
       switch (dtype.bits) {
@@ -283,73 +269,6 @@ ScalarType toScalarType(const DLDataType& dtype) {
               false, "Unsupported kDLBool bits ", std::to_string(dtype.bits));
       }
       break;
-    case DLDataTypeCode::kDLFloat8_e5m2:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e5m2;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e5m2 bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e5m2fnuz:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e5m2fnuz;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e5m2fnuz bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e4m3fn:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e4m3fn;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e4m3fn bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e4m3fnuz:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e4m3fnuz;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e4m3fnuz bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat8_e8m0fnu:
-      switch (dtype.bits) {
-        case 8:
-          stype = ScalarType::Float8_e8m0fnu;
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat8_e8m0fnu bits ", std::to_string(dtype.bits));
-      }
-      break;
-    case DLDataTypeCode::kDLFloat4_e2m1fn:
-      switch (dtype.bits) {
-        case 4:
-          switch (dtype.lanes) {
-            case 2:
-              stype = ScalarType::Float4_e2m1fn_x2;
-              break;
-            default:
-              TORCH_CHECK_BUFFER(
-                false, "Unsupported kDLFloat4_e2m1fn lanes ", std::to_string(dtype.lanes));
-          }
-          break;
-        default:
-          TORCH_CHECK_BUFFER(
-              false, "Unsupported kDLFloat4_e2m1fn bits ", std::to_string(dtype.bits));
-      }
-      break;
     default:
       TORCH_CHECK_BUFFER(false, "Unsupported code ", std::to_string(dtype.code));
   }
@@ -435,8 +354,8 @@ T* toDLPackImpl(const Tensor& src) {
   atDLMTensor->tensor.dl_tensor.device = torchDeviceToDLDevice(src.device());
   atDLMTensor->tensor.dl_tensor.ndim = static_cast<int32_t>(src.dim());
   atDLMTensor->tensor.dl_tensor.dtype = getDLDataType(src);
-  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(view.sizes().data());
-  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(view.strides().data());
+  atDLMTensor->tensor.dl_tensor.shape = view.sizes().data();
+  atDLMTensor->tensor.dl_tensor.strides = view.strides().data();
   atDLMTensor->tensor.dl_tensor.byte_offset = 0;
   fillVersion(&atDLMTensor->tensor);
 

aten/src/ATen/FunctionalStorageImpl.cpp

@@ -102,7 +102,7 @@ FunctionalStorageImpl::FunctionalStorageImpl(const Tensor& base)
   // SparseTensorImpl has no storage, so we cannot query its nbytes.
   // (original_storage_size is only used for storage resizing in fsdp anyway, which does not apply to sparse)
   // Same for XLA
-  if (base.unsafeGetTensorImpl()->has_storage() && data_ptr().device().type() != c10::DeviceType::XLA) {
+  if (base.unsafeGetTensorImpl()->has_storage() && base.device().type() != c10::DeviceType::XLA) {
     original_storage_size_ = base.unsafeGetTensorImpl()->unsafe_storage().unsafeGetStorageImpl()->sym_nbytes();
   } else {
     original_storage_size_ = -1;

aten/src/ATen/SparseTensorImpl.h

@@ -133,12 +133,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
         "resize_ called on tensor with symbolic shape")
     TORCH_CHECK(
         sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
-        "'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
-        size.size(),
-        ", sparse_dim = ",
+        "number of dimensions must be sparse_dim (",
         sparse_dim,
-        ", dense_dim = ",
-        dense_dim);
+        ") + dense_dim (",
+        dense_dim,
+        "), but got ",
+        size.size());
     if (nnz() > 0) {
       [[maybe_unused]] auto constexpr alt_options_msg =
           "You could try the following options:\n\
@@ -254,12 +254,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
         "resize_and_clear_ called on tensor with symbolic shape")
     TORCH_CHECK(
         sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
-        "'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
-        size.size(),
-        ", sparse_dim = ",
+        "number of dimensions must be sparse_dim (",
         sparse_dim,
-        ", dense_dim = ",
-        dense_dim);
+        ") + dense_dim (",
+        dense_dim,
+        "), but got ",
+        size.size());
 
     set_sizes_and_strides(size, std::vector<int64_t>(size.size()));
     sparse_dim_ = sparse_dim;

aten/src/ATen/cuda/CUDABlas.cpp

@@ -644,8 +644,6 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
   void * beta_ptr = &fbeta;
 #ifdef USE_ROCM
   int flag = 0;
-  rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
-  rocblas_datatype d_type = c_type;
 #if USE_GEMM_FLAGS_FP16_ALT_IMPL
   flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
 #endif
@@ -654,8 +652,8 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
                                    hipOperationToRocOperation(opb), (int)m, (int)n, (int)k,
                                    (void*)alpha_ptr, a, rocblas_datatype_f16_r, (int)lda, stridea,
                                    b, rocblas_datatype_f16_r, (int)ldb, strideb,
-                                   (void*)beta_ptr, c, c_type, (int)ldc, stridec,
-                                   c, d_type, (int)ldc, stridec,
+                                   (void*)beta_ptr, c, rocblas_datatype_f16_r, (int)ldc, stridec,
+                                   c, rocblas_datatype_f16_r, (int)ldc, stridec,
                                    (int) num_batches, rocblas_datatype_f32_r, rocblas_gemm_algo_standard,
                                    0, flag)));
 #else
@@ -1098,8 +1096,6 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
   GEMM_CHECK_ARGVALUES(at::Half);
 #ifdef USE_ROCM
   int flag = 0;
-  rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
-  rocblas_datatype d_type = c_type;
 #if USE_GEMM_FLAGS_FP16_ALT_IMPL
   flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
 #endif
@@ -1119,10 +1115,10 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
       ldb,
       beta_ptr,
       c,
-      c_type,
+      rocblas_datatype_f16_r,
       ldc,
       c,
-      d_type,
+      rocblas_datatype_f16_r,
       ldc,
       rocblas_datatype_f32_r,
       rocblas_gemm_algo_standard,

aten/src/ATen/cuda/CUDAGeneratorImpl.cpp

@@ -266,14 +266,11 @@ CUDAGeneratorImpl::CUDAGeneratorImpl(
  * See Note [Acquire lock when using random generators]
  */
 void CUDAGeneratorImpl::set_current_seed(uint64_t seed) {
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    state_->seed_ = seed;
-    state_->philox_offset_per_thread_ = 0;
-    no_reset_rnn_state_.clear();
-  } else {
-    TORCH_CHECK(state_->seed_ == seed, "CUDAGeneratorImpl::set_current_seed can be called during stream capture only if new seed is the same as the original seed.");
-    // no-op case
-  }
+  at::cuda::assertNotCapturing(
+      "Cannot call CUDAGeneratorImpl::set_current_seed");
+  state_->seed_ = seed;
+  state_->philox_offset_per_thread_ = 0;
+  no_reset_rnn_state_.clear();
 }
 
 /**
@@ -302,6 +299,9 @@ uint64_t CUDAGeneratorImpl::get_offset() const {
  * Gets the current seed of CUDAGeneratorImpl.
  */
 uint64_t CUDAGeneratorImpl::current_seed() const {
+  // Debatable if current_seed() should be allowed in captured regions.
+  // Conservatively disallow it for now.
+  at::cuda::assertNotCapturing("Cannot call CUDAGeneratorImpl::current_seed");
   return state_->seed_;
 }
 
@@ -346,6 +346,8 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
  * and size of the internal state.
  */
 void CUDAGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
+  at::cuda::assertNotCapturing(
+      "Please ensure to utilize the CUDAGeneratorImpl::set_state_index method during capturing.");
   static const size_t seed_size = sizeof(uint64_t);
   static const size_t offset_size = sizeof(int64_t);
   static const size_t total_size = seed_size + offset_size;
@@ -400,27 +402,15 @@ c10::intrusive_ptr<c10::GeneratorImpl> CUDAGeneratorImpl::graphsafe_get_state()
  */
 void CUDAGeneratorImpl::set_philox_offset_per_thread(uint64_t offset) {
   // see Note [Why enforce RNG offset % 4 == 0?]
-
-  // Note: If you use CUDNN RNN's, calling
-  // set_philox_offset_per_thread instead of set_offset will cause the
-  // cudnn RNN rng state to become stale.
   TORCH_CHECK(offset % 4 == 0, "offset must be a multiple of 4");
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    state_->philox_offset_per_thread_ = offset;
-  } else {
-    state_->offset_intragraph_ = offset;
-  }
+  state_->philox_offset_per_thread_ = offset;
 }
 
 /**
  * Gets the current philox_offset_per_thread_ of CUDAGeneratorImpl.
  */
 uint64_t CUDAGeneratorImpl::philox_offset_per_thread() const {
-  if (C10_LIKELY(at::cuda::currentStreamCaptureStatus() == at::cuda::CaptureStatus::None)) {
-    return state_->philox_offset_per_thread_;
-  } else {
-    return state_->offset_intragraph_;
-  }
+  return state_->philox_offset_per_thread_;
 }
 
 /**

aten/src/ATen/cuda/detail/OffsetCalculator.cuh

@@ -45,24 +45,6 @@ struct OffsetCalculator {
 
   C10_HOST_DEVICE offset_type get(index_t linear_idx) const {
     offset_type offsets;
-
-#if defined(USE_ROCM)
-    if ((dims > 0) && (dims <= 2)) {
-      auto divmod = sizes_[0].divmod(linear_idx);
-#pragma unroll
-      for (int arg = 0; arg < NARGS; arg++)
-        offsets[arg] = divmod.mod * strides_[0][arg];
-      if (dims >= 2) {
-        divmod = sizes_[1].divmod(divmod.div);
-#pragma unroll
-        for (int arg = 0; arg < NARGS; arg++)
-          offsets[arg] += divmod.mod * strides_[1][arg];
-      }
-      // [...]
-      return offsets;
-    }
-#endif
-
     #pragma unroll
     for (int arg = 0; arg < NARGS; arg++) {
       offsets[arg] = 0;

aten/src/ATen/dlpack.h

@@ -19,7 +19,7 @@
 #define DLPACK_MAJOR_VERSION 1
 
 /*! \brief The current minor version of dlpack */
-#define DLPACK_MINOR_VERSION 1
+#define DLPACK_MINOR_VERSION 0
 
 /*! \brief DLPACK_DLL prefix for windows */
 #ifdef _WIN32
@@ -32,7 +32,9 @@
 #define DLPACK_DLL
 #endif
 
+// NOLINTNEXTLINE(modernize-deprecated-headers)
 #include <stdint.h>
+// NOLINTNEXTLINE(modernize-deprecated-headers)
 #include <stddef.h>
 
 #ifdef __cplusplus
@@ -157,26 +159,6 @@ typedef enum {
   kDLComplex = 5U,
   /*! \brief boolean */
   kDLBool = 6U,
-  /*! \brief FP8 data types */
-  kDLFloat8_e3m4 = 7U,
-  kDLFloat8_e4m3 = 8U,
-  kDLFloat8_e4m3b11fnuz = 9U,
-  kDLFloat8_e4m3fn = 10U,
-  kDLFloat8_e4m3fnuz = 11U,
-  kDLFloat8_e5m2 = 12U,
-  kDLFloat8_e5m2fnuz = 13U,
-  kDLFloat8_e8m0fnu = 14U,
-  /*! \brief FP6 data types
-   * Setting bits != 6 is currently unspecified, and the producer must ensure it is set
-   * while the consumer must stop importing if the value is unexpected.
-   */
-  kDLFloat6_e2m3fn = 15U,
-  kDLFloat6_e3m2fn = 16U,
-  /*! \brief FP4 data types
-   * Setting bits != 4 is currently unspecified, and the producer must ensure it is set
-   * while the consumer must stop importing if the value is unexpected.
-   */
-  kDLFloat4_e2m1fn = 17U,
 } DLDataTypeCode;
 
 /*!
@@ -190,12 +172,6 @@ typedef enum {
  *   - int8: type_code = 0, bits = 8, lanes = 1
  *   - std::complex<float>: type_code = 5, bits = 64, lanes = 1
  *   - bool: type_code = 6, bits = 8, lanes = 1 (as per common array library convention, the underlying storage size of bool is 8 bits)
- *   - float8_e4m3: type_code = 8, bits = 8, lanes = 1 (packed in memory)
- *   - float6_e3m2fn: type_code = 16, bits = 6, lanes = 1 (packed in memory)
- *   - float4_e2m1fn: type_code = 17, bits = 4, lanes = 1 (packed in memory)
- *
- *  When a sub-byte type is packed, DLPack requires the data to be in little bit-endian, i.e.,
- *  for a packed data set D ((D >> (i * bits)) && bit_mask) stores the i-th element.
  */
 typedef struct {
   /*!
@@ -253,12 +229,12 @@ typedef struct {
   /*! \brief The data type of the pointer*/
   DLDataType dtype;
   /*! \brief The shape of the tensor */
-  int64_t* shape;
+  const int64_t* shape;
   /*!
    * \brief strides of the tensor (in number of elements, not bytes)
    *  can be NULL, indicating tensor is compact and row-majored.
    */
-  int64_t* strides;
+  const int64_t* strides;
   /*! \brief The offset in bytes to the beginning pointer to data */
   uint64_t byte_offset;
 } DLTensor;
@@ -293,7 +269,7 @@ typedef struct DLManagedTensor {
   void (*deleter)(struct DLManagedTensor * self);
 } DLManagedTensor;
 
-// bit masks used in the DLManagedTensorVersioned
+// bit masks used in in the DLManagedTensorVersioned
 
 /*! \brief bit mask to indicate that the tensor is read only. */
 #define DLPACK_FLAG_BITMASK_READ_ONLY (1UL << 0UL)
@@ -306,14 +282,6 @@ typedef struct DLManagedTensor {
  */
 #define DLPACK_FLAG_BITMASK_IS_COPIED (1UL << 1UL)
 
-/*
- * \brief bit mask to indicate that whether a sub-byte type is packed or padded.
- *
- * The default for sub-byte types (ex: fp4/fp6) is assumed packed. This flag can
- * be set by the producer to signal that a tensor of sub-byte type is padded.
- */
-#define DLPACK_FLAG_BITMASK_IS_SUBBYTE_TYPE_PADDED (1UL << 2UL)
-
 /*!
  * \brief A versioned and managed C Tensor object, manage memory of DLTensor.
  *

aten/src/ATen/functorch/BatchRulesUnaryOps.cpp

@@ -171,8 +171,6 @@ TORCH_LIBRARY_IMPL(aten, FuncTorchBatched, m) {
 
   POINTWISE_BOXED(fill_.Scalar);
   POINTWISE_BOXED(zero_);
-  // This is special because this op doesn't return anything
-  m.impl("_assert_tensor_metadata", native::_assert_tensor_metadata);
 
 #undef UNARY_POINTWISE
 #undef UNARY_POINTWISE_ALL

aten/src/ATen/native/CPUBlas.cpp

@@ -457,9 +457,24 @@ void gemm(
     return;
   }
 #endif
+  // for the fallback path, first compute gemm with beta = 0,
+  // and then add c in full precision.
+  int64_t c_size = n * m;
+  std::vector<float> float_c(c_size, 0.f);
   gemm_no_downcast_stub(
       at::kCPU, at::kBFloat16,
-      transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+      transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
+  for (const auto j : c10::irange(n)) {
+    for (const auto i : c10::irange(m)) {
+      auto offset = j * ldc + i;
+      // beta == 0 won't propagate NaN from C
+      if (beta == 0.f) {
+        c[offset] = float_c[j * m + i];
+      } else {
+        c[offset] = beta * c[offset] + float_c[j * m + i];
+      }
+    }
+  }
 }
 
 void gemm(
@@ -478,9 +493,24 @@ void gemm(
     return;
   }
 #endif
+  // for the fallback path, first compute gemm with beta = 0,
+  // and then add c in full precision.
+  int64_t c_size = n * m;
+  std::vector<float> float_c(c_size, 0.f);
   gemm_no_downcast_stub(
       at::kCPU, at::kHalf,
-      transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+      transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
+  for (const auto j : c10::irange(n)) {
+    for (const auto i : c10::irange(m)) {
+      auto offset = j * ldc + i;
+      // beta == 0 won't propagate NaN from C
+      if (beta == 0.f) {
+        c[offset] = float_c[j * m + i];
+      } else {
+        c[offset] = beta * c[offset] + float_c[j * m + i];
+      }
+    }
+  }
 }
 
 void gemm(

aten/src/ATen/native/ChanelShuffle.cpp

@@ -81,7 +81,7 @@ Tensor math_channel_shuffle(const Tensor& self, int64_t groups) {
   // TODO: contiguous can be made to preserve the memory format
   // of the input. However since the above reshape clobbers h and w
   // it may not be safe to do that, since channels_last contiguous
-  // may think oc and the last dim correspond to h,w?
+  // may think oc and and the last dim correspond to h,w?
   // It is not clear, however from initial looking around it feels that
   // this may not be correct.
   // In this case channels last will likely require custom implementation

aten/src/ATen/native/EmbeddingBag.h

@@ -1,4 +1,3 @@
-#pragma once
 #include <ATen/core/Tensor.h>
 #include <ATen/Config.h>
 #include <cstdint>

aten/src/ATen/native/FractionalMaxPool3d.cpp

@@ -67,13 +67,13 @@ TORCH_PRECOMPUTE_META_FUNC(fractional_max_pool3d)(
   int64_t inputH = input_.size(heightDim);
   int64_t inputW = input_.size(widthDim);
 
-  TORCH_CHECK((poolSizeT <= inputT) && (outputT + poolSizeT - 1 < inputT),
+  TORCH_CHECK(outputT + poolSizeT - 1 < inputT,
            "fractional_max_pool3d_out(): pool time ", poolSizeT,
            " too large relative to input time ", inputT);
-  TORCH_CHECK((poolSizeW <= inputW) && (outputW + poolSizeW - 1 < inputW),
+  TORCH_CHECK(outputW + poolSizeW - 1 < inputW,
            "fractional_max_pool3d_out(): pool width ", poolSizeW,
            " too large relative to input width ", inputW);
-  TORCH_CHECK((poolSizeH <= inputH) && (outputH + poolSizeH - 1 < inputH),
+  TORCH_CHECK(outputH + poolSizeH - 1 < inputH,
            "fractional_max_pool3d_out(): pool height ", poolSizeH,
            " too large relative to input height ", inputH);
 

aten/src/ATen/native/LinearAlgebra.cpp

@@ -1360,8 +1360,7 @@ Tensor outer(const Tensor& self, const Tensor& vec2) {
 #endif
 
 
-#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
-// Used by default on x86 platforms and on AArch64+ACL
+#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
 static inline int64_t get_mkldnn_matmul_min_dim() {
   static auto value = [&] {
     const int64_t default_min_dim = [&] {
@@ -1396,6 +1395,8 @@ static inline bool apply_mkldnn_matmul_heur(int64_t m, int64_t k, int64_t n) {
   return at::globalContext().userEnabledMkldnn() && m > min_dim && k > min_dim && n > min_dim && m * k * n > min_size;
 }
 #endif
+
+
 static void addmm_impl_cpu_(
     Tensor &result, const Tensor &self, Tensor m1, Tensor m2, const Scalar& beta, const Scalar& alpha) {
   TORCH_INTERNAL_ASSERT(self.dim() == 2 && m1.dim() == 2 && m2.dim() == 2);
@@ -1771,8 +1772,8 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
     return (strides[2] == 1 && (sizes[1] == 1 || strides[1] >= sizes[2])) ||
         (strides[1] == 1 && (sizes[2] == 1 || strides[2] >= sizes[1]));
   };
-#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
-  // Always apply mkldnn heuristic on x86 platform, but on ARM only if compiled with ACL
+
+#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
   bool apply_heur = apply_mkldnn_matmul_heur(batch1.sizes()[1], batch1.sizes()[2], batch2.sizes()[2]);
   if (apply_heur && use_mkldnn_matmul(batch1, batch2, self_or_result)) {
     try {
@@ -1784,6 +1785,7 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
     }
   }
 #endif
+
   if (contraction_size * res_rows * res_cols < 400) {
     if (is_bmm_out) {
       AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, batch1.scalar_type(), "bmm", [&] {

aten/src/ATen/native/Normalization.cpp

@@ -624,9 +624,7 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, int64_t> _batch_norm_impl_index(
   if (backend == BatchNormBackend::Miopen) {
     return std::tuple_cat(
              at::miopen_batch_norm(
-               input.contiguous(input.suggest_memory_format()),
-               weight.contiguous(),
-               bias.contiguous(),
+               input.contiguous(), weight.contiguous(), bias.contiguous(),
                running_mean.defined() ? running_mean.contiguous() : running_mean,
                running_var.defined() ? running_var.contiguous() : running_var,
                training, momentum, eps),

aten/src/ATen/native/TensorAdvancedIndexing.cpp

@@ -2174,7 +2174,7 @@ static void _scatter_via_index_put(
   if (self.dim() == 1 || broadcast_index) {
     Tensor squeezed = index;
     if (broadcast_index && index.dim() > 1) {
-      for (int64_t d = index.dim() - 1; d >= 0; --d) {
+      for (const auto d : c10::irange(index.dim())) {
         if (d == dim) {
           continue;
         }

aten/src/ATen/native/TriangularOps.cpp

@@ -52,7 +52,6 @@ void apply_triu_tril_single(
     int64_t self_col_stride,
     bool upper) {
   constexpr int64_t zero = 0;
-  k = std::clamp(k, -n, m); // Clamp k to [-n, m] to prevent i + k arithmetic overflow, especially if k approaches INT64_MAX/INT64_MIN.
 
   if (upper) {
     parallel_for(0, n, 0, [&](int64_t start, int64_t end) {

aten/src/ATen/native/cpu/MaxUnpoolKernel.cpp

@@ -85,11 +85,11 @@ void cpu_max_unpool(
     if constexpr (is_3d) {
       TORCH_CHECK(false, "Found an invalid max index: ", optional_error_index.value(),
           " (output volumes are of size ", output_depth,
-          "x", output_height, "x", output_width, ")");
+          "x", output_height, "x", output_width);
     } else {
       TORCH_CHECK(false, "Found an invalid max index: ", optional_error_index.value(),
           " (output volumes are of size ", output_height,
-          "x", output_width, ")");
+          "x", output_width);
     }
   }
 

aten/src/ATen/native/cuda/ActivationHardsigmoidKernel.cu

@@ -36,7 +36,7 @@ void hardsigmoid_kernel(TensorIteratorBase& iter) {
             [zero, one_sixth, three, six] GPU_LAMBDA(
                 scalar_t self_val) -> scalar_t {
               opmath_t x = static_cast<opmath_t>(self_val);
-              return std::min<opmath_t>(std::max<opmath_t>(x + three, zero), six) * one_sixth;
+              return std::min(std::max(x + three, zero), six) * one_sixth;
             });
       });
 }

aten/src/ATen/native/cuda/Blas.cpp

@@ -1080,6 +1080,16 @@ static bool _scaled_mm_allowed_device(bool sm90_only=false, bool sm100_only=fals
 #endif
 }
 
+static bool _grouped_mm_allowed_device() {
+#ifdef USE_ROCM
+    return false;
+#else
+    auto dprops = at::cuda::getCurrentDeviceProperties();
+    // CUDA capability 8.0 and greater
+    return dprops->major >= 8;
+#endif
+}
+
 #ifdef USE_ROCM
 static bool _scaled_mm_is_fnuz() {
     return at::detail::getCUDAHooks().isGPUArch({"gfx942"});
@@ -1776,19 +1786,14 @@ Tensor _grouped_mm_cuda(const Tensor& mat_a, const Tensor& mat_b,
 const std::optional<at::Tensor>& offs,
 const std::optional<at::Tensor>& bias,
 std::optional<c10::ScalarType> out_dtype) {
+#ifndef USE_ROCM
   _grouped_mm_validate_inputs(mat_a, mat_b, offs, bias, out_dtype);
   bool a_b_and_out_are_bf16 = (
     mat_a.dtype() == at::kBFloat16 &&
     mat_b.dtype() == at::kBFloat16 &&
     out_dtype.value_or(at::kBFloat16) == at::kBFloat16
   );
-#ifndef USE_ROCM
   bool use_fast_path = _scaled_mm_allowed_device(/*sm90_only*/true, /*sm100_only*/true) && a_b_and_out_are_bf16;
-#else
-  // _scaled_mm_allowed_device is used here within _grouped_mm_cuda which seems incorrect since scale is not used.
-  // the _grouped_mm_fallback should be safe for any ROCm GPU since it's just calling typical mm/bmm
-  bool use_fast_path = false;
-#endif
   const auto out_dtype_ = _resolve_grouped_mm_out_dtype(mat_a, mat_b, out_dtype);
   Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
   if (use_fast_path) {
@@ -1798,6 +1803,9 @@ std::optional<c10::ScalarType> out_dtype) {
     _grouped_mm_fallback(mat_a, mat_b, offs, bias, out_dtype, out);
   }
   return out;
+#else
+  TORCH_CHECK(false, "grouped gemm is not supported on ROCM")
+#endif
 }
 
 Tensor _bmm_dtype_cuda(const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype) {

aten/src/ATen/native/cuda/Nonzero.cu

@@ -317,17 +317,6 @@ void nonzero_static_cuda_out_impl(
     out_temp =
         Tensor(at::detail::empty_cuda({self.dim(), size}, out.options())).t();
   }
-  // If input has zero elements, avoid kernel grid calculations (which can
-  // produce zero divisors) and just fill the output with fill_value.
-  if (self.numel() == 0) {
-    if (need_to_copy) {
-      out_temp.fill_(fill_value);
-      out.copy_(out_temp);
-    } else {
-      out.fill_(fill_value);
-    }
-    return;
-  }
   int64_t* out_data_ptr = need_to_copy ? out_temp.mutable_data_ptr<int64_t>()
                                        : out.mutable_data_ptr<int64_t>();
 

aten/src/ATen/native/cuda/Reduce.cuh

@@ -416,7 +416,6 @@ struct ReduceOp {
     if (config.should_block_y_reduce()) {
       value = block_y_reduce<output_vec_size>(value, shared_memory);
     }
-    __syncthreads();
     if (config.should_block_x_reduce()) {
       value = block_x_reduce<output_vec_size>(value, shared_memory);
     }

aten/src/ATen/native/cuda/Repeat.cu

@@ -17,11 +17,12 @@ __global__ static void compute_cuda_kernel(
     index_t* result_ptr,
     int64_t size,
     int64_t result_size) {
-  CUDA_KERNEL_ASSERT_PRINTF(
-      result_size == cumsum_ptr[size - 1],
+  if (C10_UNLIKELY((result_size != cumsum_ptr[size - 1]))) {
+    printf("%s:%d:%s: block: [%d,%d,%d], thread: [%d,%d,%d] "
       "Invalid input! In `repeat_interleave`, the `output_size` argument (%ld) must be the same as the sum of the elements in the `repeats` tensor (%ld).\n",
-      result_size,
-      cumsum_ptr[size - 1]);
+      __FILE__, __LINE__, __func__,blockIdx.x, blockIdx.y, blockIdx.z, threadIdx.x, threadIdx.y, threadIdx.z, result_size, cumsum_ptr[size - 1 ]);
+    CUDA_KERNEL_ASSERT(result_size == cumsum_ptr[size - 1])
+  }
 
   int64_t idx = ((int64_t) blockIdx.x) * blockDim.x + threadIdx.x;
   int64_t stride = (blockDim.x * gridDim.x) / C10_WARP_SIZE;

aten/src/ATen/native/cuda/Shape.cu

@@ -226,38 +226,6 @@ __global__ void CatArrayBatchedCopy_contig(
     }
 }
 
-
-template <typename T, typename IndexType, int Dims, int batch_size, int stride_size, int alignment, int elems_per_vec>
-__global__ void CatArrayBatchedCopy_vectorized(
-    char* output,
-    CatArrInputTensorMetadata<T, IndexType, batch_size, stride_size> inputs,
-    TensorSizeStride<IndexType, CAT_ARRAY_MAX_INPUT_DIMS> os,
-    const int concatDim,
-    IndexType trailingSize) {
-
-    IndexType tid = blockIdx.x * blockDim.x + threadIdx.x;
-    IndexType nElements = inputs.nElements[blockIdx.y] / elems_per_vec;
-
-    if(tid >= nElements) return;
-
-    const char * data = (char*)inputs.input[blockIdx.y];
-    IndexType offset = inputs.offset[blockIdx.y] * trailingSize / elems_per_vec;
-    IndexType dimSize = inputs.dimSize[blockIdx.y] * trailingSize / elems_per_vec;
-    int64_t dataOffset = (int64_t)offset  * alignment; // in bytes
-
-    IndexType stride = gridDim.x * blockDim.x;
-
-    while( tid < nElements){
-      int64_t elementOffset = (int64_t)CatArrIndexToOffset<IndexType, Dims>::compute(
-                    os.tensorSize, os.tensorStride, dimSize, concatDim, tid) * alignment; // in bytes
-      auto vec = at::native::memory::ld_vec<alignment>(data + (int64_t)alignment * tid);
-      at::native::memory::st_vec<alignment>(output + dataOffset + elementOffset, vec);
-      tid += stride;
-    }
-}
-
-
-
 /*
   Specialized implementation of the CatArrayBatchedCopy written to generate wide memory loads
   to improve memory bandwidth throughput.
@@ -328,27 +296,12 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
   scalar_t *data = (scalar_t *)(out.mutable_data_ptr());
   CatArrInputTensorMetadata<scalar_t, unsigned int, batch_size, stride_size> catMetaData;
   TensorSizeStride<unsigned int, CAT_ARRAY_MAX_INPUT_DIMS> outputParam;
-  // If all batches are contiguous we can call a specialized implementation
-  // which requires the input tensor addresses to be aligned to a
-  // 16 Byte boundary.
-
-  constexpr bool isContig = stride_size == 1;
-  bool isAligned = true;
-  constexpr int alignment = 16;
 
   // Next, let's initialize the size, stride arrays for the output Tensor.
-  // for contig case, we'll canonicalize output strides, so that
-  // we don't have arbitrary strides for dims of size 0
-  size_t stride0 = 1;
   if (memory_format == c10::MemoryFormat::Contiguous) {
-    for (int i = nDims - 1; i >= 0; --i) {
+    for (int i = 0; i < nDims; ++i) {
       outputParam.tensorSize[i] = out.size(i);
-      if (isContig) {
-        outputParam.tensorStride[i] = stride0;
-        stride0 *= out.size(i);
-      } else {
-        outputParam.tensorStride[i] = out.stride(i);
-      }
+      outputParam.tensorStride[i] = out.stride(i);
     }
   } else if (memory_format == c10::MemoryFormat::ChannelsLast || memory_format == c10::MemoryFormat::ChannelsLast3d) {
     // permute the semantics of dims from NCHW to NHWC so that the input
@@ -367,15 +320,12 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
 
   at::cuda::CUDAStream stream = at::cuda::getCurrentCUDAStream();
 
+  // If all batches are contiguous we can call a specialized implementation
+  // which requires the input tensor addresses to be aligned to a
+  // 16 Byte boundary.
 
-  // for channels last computing slice size correctly is much more involved, so we never send it
-  // on the fully vectorized path
-  // we need output stride in cat dimension to be multiple of alignment,
-  // if we ever use it to compute offsets
-  // for catting in 0th dimension it doesn't matter
-  bool isInOutAligned = isContig && at::native::memory::get_alignment(data) >= alignment &&
-                        memory_format == c10::MemoryFormat::Contiguous && (dimension == 0 ||
-                        outputParam.tensorStride[dimension - 1] * sizeof(scalar_t) % alignment == 0);
+  bool isContig = true;
+  bool isAligned = true;
   unsigned int max_elements_per_tensor = 0;
 
   // Now we loop
@@ -391,16 +341,6 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
       // high-dimensional tensor
       if (inputs[i+batchCounter].get().numel() > 0) {
         dimSize = inputs[i+batchCounter].get().size(dimension);
-        if (isInOutAligned) {
-          auto t = inputs[i+batchCounter].get();
-          // similarly to output stride, we cannot trust stride value to
-          // determine slice size if the corresponding dimension is 1
-          // we have to multiply all the subsequent sizes
-          int64_t slice_size = dimension == 0 ? t.numel() : t.sizes()[dimension - 1] != 1 ?
-             t.strides()[dimension - 1] : c10::multiply_integers(t.sizes().begin() + dimension, t.sizes().end());
-          slice_size *= sizeof(scalar_t);
-          isInOutAligned &= (slice_size % alignment == 0);
-        }
       }
 
       catMetaData.input[batchCounter] = (scalar_t*)(inputs[i+batchCounter].get().const_data_ptr());
@@ -411,12 +351,10 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
 #ifdef USE_ROCM
       // On ROCm, CatArrayBatchedCopy_contig is faster
       isAligned = false;
-      isInOutAligned = false;
 #else
       // If at least one of the inputs is not aligned, we can't call the
       // CatArrayBatchedCopy_alignedK_contig
       isAligned &= is_aligned_vec4(catMetaData.input[batchCounter]);
-      isInOutAligned &= at::native::memory::get_alignment(catMetaData.input[batchCounter]) >= alignment;
 #endif
 
       if (stride_size > 1) {
@@ -427,6 +365,7 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
           catMetaData.tensorStride[batchCounter].tensorStride[j] = strides[j];
         }
         catMetaData.isContiguous[batchCounter] = false;
+        isContig = false;
       } else {
         catMetaData.isContiguous[batchCounter] = true;
       }
@@ -449,13 +388,10 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
           max_elements_per_tensor, batchCounter);
 #else
     dim3 applyBlock, catGrid;
-    if (isInOutAligned) {
-      std::tie(catGrid, applyBlock) = getCatGridContig<scalar_t, alignment>(
-        max_elements_per_tensor, batchCounter);
-    } else if (isContig && isAligned && sizeof(scalar_t) > 2) {
+    if (isContig && sizeof(scalar_t) > 2) {
       std::tie(catGrid, applyBlock) = getCatGridContig<scalar_t, ALIGNED_VEC_LOAD_BYTES_16>(
           max_elements_per_tensor, batchCounter);
-    } else if (isContig && isAligned && sizeof(scalar_t) == 2) {
+    } else if (isContig && sizeof(scalar_t) == 2) {
       std::tie(catGrid, applyBlock) = getCatGridContig<scalar_t, ALIGNED_VEC_LOAD_BYTES_8>(
           max_elements_per_tensor, batchCounter);
     } else {
@@ -463,30 +399,6 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
       getCatGrid(batchCounter, catGrid);
     }
 #endif
-    int32_t trailingSize;
-    TensorSizeStride<unsigned int, CAT_ARRAY_MAX_INPUT_DIMS> kernelOutputParam;
-    if (isInOutAligned) {
-      // in this case we can and should flatten the tensors after the cat dim
-      // we want to view the tensors as if consisting of `alignment`-sized elements
-      // however, we might not be able to cleanly divide just the last dim -
-      // it might not be the multiple of alignment.
-      // however, we know that the full concatted slice is multiple of alignment,
-      // so if we flatten all the dims after and including concat dim,
-      // it will be divisible by alignment
-      // then we need to divide last out size by elems_per_vec,
-      // and divide all strides except last by elems_per_vec (last stride is 1 always)
-      // for input, we will fix up the sizes and strides in the kernel directly
-      kernelOutputParam = outputParam;
-      nDims = dimension + 1;
-      constexpr auto elems_per_vec = alignment / sizeof(scalar_t);
-      auto out_size = dimension == 0 ? out.numel() : kernelOutputParam.tensorStride[dimension-1];
-      kernelOutputParam.tensorSize[dimension] = out_size / elems_per_vec;
-      trailingSize = outputParam.tensorStride[dimension];
-      kernelOutputParam.tensorStride[dimension] = 1;
-      for (int i = 0; i < dimension; ++i) {
-        kernelOutputParam.tensorStride[i] /= elems_per_vec;
-      }
-    }
 
     if (memory_format != c10::MemoryFormat::Contiguous) {
       switch (dimension) {
@@ -501,12 +413,7 @@ void parallel_cat(const Tensor &out, const MaterializedITensorListRef& inputs, i
     }
     // Template Declarations for dim = 1, 2, 3, 4
 #define HANDLE_CASE(DIMS) \
-    if (isInOutAligned) {\
-      constexpr auto elems_per_vec = alignment / sizeof(scalar_t); \
-      CatArrayBatchedCopy_vectorized<scalar_t, unsigned int, DIMS, batch_size, stride_size, alignment, elems_per_vec><<<\
-      catGrid, applyBlock, 0, stream.stream()>>>(\
-        (char*)data, catMetaData, kernelOutputParam, dimension, trailingSize);\
-    } else if (isContig && isAligned && sizeof(scalar_t) > 2 && sizeof(scalar_t) <= 8) {\
+    if (isContig && isAligned && sizeof(scalar_t) > 2 && sizeof(scalar_t) <= 8) {\
       CatArrayBatchedCopy_alignedK_contig<scalar_t, unsigned int, DIMS, batch_size, stride_size, ALIGNED_VEC_LOAD_BYTES_16><<<\
           catGrid, applyBlock, 0, stream.stream()>>>(\
               data, catMetaData, outputParam, dimension, outputParam.tensorStride[dimension]);\

aten/src/ATen/native/cuda/int8mm.cu

@@ -5,20 +5,12 @@
 
 namespace at::native {
 
-__global__ void weight_int8pack_mm_kernel(
-    const float* x,
-    const int8_t* w,
-    const float* scale,
-    float* out,
-    int B,
-    int K,
-    int N) {
+__global__ void weight_int8pack_mm_kernel(const float* x, const int8_t* w, const float* scale, float* out, int B, int K, int N) {
   // one thread per output element: [B, N]
   int b = blockIdx.y * blockDim.y + threadIdx.y;
   int n = blockIdx.x * blockDim.x + threadIdx.x;
 
-  if (b >= B || n >= N)
-    return;
+  if (b >= B || n >= N) return;
 
   float acc = 0.0f;
   for (int k = 0; k < K; ++k) {
@@ -28,11 +20,7 @@ __global__ void weight_int8pack_mm_kernel(
   out[b * N + n] = acc * scale[n];
 }
 
-void launch_weight_int8pack_mm_cuda_kernel(
-    const Tensor& x,
-    const Tensor& w_int8,
-    const Tensor& scale,
-    Tensor& out) {
+void launch_weight_int8pack_mm_cuda_kernel(const Tensor& x, const Tensor& w_int8, const Tensor& scale, Tensor& out) {
   const int B = x.size(0);
   const int K = x.size(1);
   const int N = w_int8.size(0);
@@ -47,16 +35,12 @@ void launch_weight_int8pack_mm_cuda_kernel(
       w_int8.data_ptr<int8_t>(),
       scale.data_ptr<float>(),
       out.data_ptr<float>(),
-      B,
-      K,
-      N);
+      B, K, N);
 }
 
+
 // Main GPU entry point
-at::Tensor _weight_int8pack_mm_cuda(
-    const at::Tensor& x,
-    const at::Tensor& w_int8,
-    const at::Tensor& scale) {
+at::Tensor _weight_int8pack_mm_cuda(const at::Tensor& x, const at::Tensor& w_int8, const at::Tensor& scale) {
   // --- Check inputs ---
   TORCH_CHECK(x.is_cuda(), "x must be a CUDA tensor");
   TORCH_CHECK(w_int8.is_cuda(), "w must be a CUDA tensor");
@@ -66,16 +50,12 @@ at::Tensor _weight_int8pack_mm_cuda(
   TORCH_CHECK(w_int8.dim() == 2, "w must be 2D");
   TORCH_CHECK(scale.dim() == 1, "scale must be 1D");
 
-  TORCH_CHECK(
-      x.size(1) == w_int8.size(1),
-      "K dimension mismatch: x.size(1) != w.size(1)");
-  TORCH_CHECK(
-      w_int8.size(0) == scale.size(0),
-      "Output dim mismatch: w.size(0) != scale.size(0)");
+  TORCH_CHECK(x.size(1) == w_int8.size(1), "K dimension mismatch: x.size(1) != w.size(1)");
+  TORCH_CHECK(w_int8.size(0) == scale.size(0), "Output dim mismatch: w.size(0) != scale.size(0)");
 
   // --- Determine shapes ---
-  auto B = x.size(0); // batch size
-  auto N = w_int8.size(0); // output dim
+  auto B = x.size(0);  // batch size
+  auto N = w_int8.size(0);  // output dim
 
   // Ensure inputs are in the correct types for the kernel
   auto x_f32 = x.to(at::kFloat);
@@ -83,13 +63,12 @@ at::Tensor _weight_int8pack_mm_cuda(
   auto scale_f32 = scale.to(at::kFloat);
 
   // --- Allocate output ---
-  auto out = at::empty({B, N}, x_f32.options());
+  auto out = at::empty({B, N}, x.options().dtype(at::kFloat));
 
   // --- Launch kernel ---
-  launch_weight_int8pack_mm_cuda_kernel(
-      x_f32, w_int8_contiguous, scale_f32, out);
+  launch_weight_int8pack_mm_cuda_kernel(x_f32, w_int8_contiguous, scale_f32, out);
 
-  return out.to(x.dtype());
+  return out;
 }
 
 } // namespace at::native

aten/src/ATen/native/cudnn/MHA.cpp

@@ -482,9 +482,7 @@ auto build_graph(
   auto scaled_dot_product_flash_attention_options =
       fe::graph::SDPA_attributes()
           .set_name("CUDNN_SDPA")
-          .set_is_inference(return_softmaxstats == false)
-          // TODO(eqy): switch to this API once cuDNN FE is upgraded
-          // .set_generate_stats(return_softmaxstats)
+          .set_generate_stats(return_softmaxstats)
           .set_causal_mask(is_causal)
           .set_attn_scale(attn_scale);
   if (use_ragged_in_dense(q, k, v, o, attn_bias.has_value())) {
@@ -704,9 +702,7 @@ auto build_graph_nestedtensor(
   auto scaled_dot_product_flash_attention_options =
       fe::graph::SDPA_attributes()
           .set_name("CUDNN_SDPA_NESTEDTENSOR")
-          .set_is_inference(return_softmaxstats == false)
-          // TODO(eqy): switch to this API once cuDNN FE is upgraded
-          // .set_generate_stats(return_softmaxstats)
+          .set_generate_stats(return_softmaxstats)
           .set_causal_mask(is_causal)
           .set_attn_scale(attn_scale)
           .set_seq_len_q(SEQ_LEN_Q_)

aten/src/ATen/native/miopen/BatchNorm_miopen.cpp

@@ -7,7 +7,6 @@
 #include <ATen/NativeFunctions.h>
 #else
 #include <ATen/ops/empty.h>
-#include <ATen/ops/empty_like.h>
 #include <ATen/ops/miopen_batch_norm_native.h>
 #include <ATen/ops/miopen_batch_norm_backward_native.h>
 #endif
@@ -103,7 +102,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm(
     mode = miopenBNSpatial;
   }
 
-  auto output_t = at::empty_like(input_t, input_t.options(), input_t.suggest_memory_format());
+  auto output_t = at::empty(input->sizes(), input->options());
   TensorArg output{ output_t, "output", 0 };
 
   auto handle = getMiopenHandle();
@@ -171,15 +170,20 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
     const std::optional<Tensor>& save_var_t_opt,
     double epsilon) {
   // See [Note: hacky wrapper removal for optional tensor]
-  const Tensor& save_mean_t = save_mean_t_opt.value_or(Tensor());
-  const Tensor& save_var_t = save_var_t_opt.value_or(Tensor());
+  const Tensor& running_mean =
+      running_mean_opt.value_or(Tensor());
+  const Tensor& running_var =
+      running_var_opt.value_or(Tensor());
+  const Tensor& save_mean_t =
+      save_mean_t_opt.value_or(Tensor());
+  const Tensor& save_var_t =
+      save_var_t_opt.value_or(Tensor());
 
-  auto grad_output_contig =
-      grad_output_t.contiguous(input_t.suggest_memory_format());
-  TensorArg input{input_t, "input", 1},
-      grad_output{grad_output_contig, "grad_output", 2},
-      weight{weight_t, "weight", 3}, save_mean{save_mean_t, "save_mean", 4},
-      save_var{save_var_t, "save_var", 5};
+  TensorArg input{ input_t, "input", 1 },
+            grad_output{ grad_output_t, "grad_output", 2 },
+            weight{ weight_t, "weight", 3 },
+            save_mean{ save_mean_t, "save_mean", 4 },
+            save_var{ save_var_t, "save_var", 5 };
   CheckedFrom c = "miopen_batch_norm_backward";
 
   checkAllDefined(c, {input, grad_output, weight, save_mean, save_var});
@@ -191,11 +195,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
   }
   checkAllSameType(c, {input, grad_output});
   checkAllSameType(c, {weight, save_mean, save_var});
-  // TODO: is weight required to be contiguous?
-  checkAllContiguous(c, {save_mean, save_var});
-  // TODO: TensorArg check should start handle memory format
-  TORCH_CHECK(input->is_contiguous(input->suggest_memory_format()));
-  TORCH_CHECK(grad_output->is_contiguous(input->suggest_memory_format()));
+  checkAllContiguous(c, {input, grad_output, save_mean, save_var});
   checkDimRange(c, input, 2, 6 /* exclusive */);
   checkSameSize(c, input, grad_output);
   auto num_features = input->size(1);
@@ -210,7 +210,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
     mode = miopenBNSpatial;
   }
 
-  auto grad_input_t  = at::empty(input->sizes(), input->options(), input->suggest_memory_format());
+  auto grad_input_t  = at::empty(input->sizes(), input->options());
   auto grad_weight_t = at::empty(weight->sizes(), weight->options());
   auto grad_bias_t   = at::empty(weight->sizes(), weight->options());
 

aten/src/ATen/native/miopen/Conv_miopen.cpp

@@ -1770,12 +1770,10 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> miopen_depthwise_convolution_back
 // fusions
 // ---------------------------------------------------------------------
 
-void raw_miopen_convolution_add_relu_out(
+void raw_miopen_convolution_relu_out(
     const Tensor& output,
     const Tensor& input,
     const Tensor& weight,
-    const Tensor& z,
-    float alpha,
     const Tensor& bias,
     IntArrayRef stride,
     IntArrayRef padding,
@@ -1783,20 +1781,68 @@ void raw_miopen_convolution_add_relu_out(
     int64_t groups,
     bool benchmark,
     bool deterministic) {
-  raw_miopen_convolution_forward_out(
-      output,
+  auto dataType = getMiopenDataType(input);
+  miopenConvolutionMode_t c_mode = miopenConvolution;
+  ConvolutionArgs args{ input, output, weight };
+  args.handle = getMiopenHandle();
+  at::MemoryFormat memory_format = miopen_conv_suggest_memory_format(input, weight);
+  setConvolutionParams(
+      &args.params,
+      args.handle,
       input,
       weight,
       padding,
       stride,
       dilation,
       groups,
+      deterministic,
+      memory_format);
+  args.idesc.set(input, memory_format);
+  args.wdesc.set(weight, memory_format, 0);
+  args.odesc.set(output, memory_format);
+  args.cdesc.set(
+      dataType,
+      c_mode,
+      input.dim() - 2,
+      args.params.padding,
+      args.params.stride,
+      args.params.dilation,
+      args.params.groups,
       benchmark,
       deterministic);
-  at::Tensor alpha_mul_z_add_bias =
-      at::native::reshape_bias(input.dim(), bias).add(z, alpha);
-  output.add_(alpha_mul_z_add_bias);
-  output.relu_();
+
+  TensorDescriptor bdesc;
+  bdesc.set(bias.expand({1, bias.size(0)}), output.dim());
+
+  // Create the fusion plan
+  miopenFusionPlanDescriptor_t fusePlanDesc;
+  miopenFusionOpDescriptor_t convoOp;
+  miopenFusionOpDescriptor_t biasOp;
+  miopenFusionOpDescriptor_t activOp;
+  MIOPEN_CHECK(miopenCreateFusionPlan(&fusePlanDesc, miopenVerticalFusion, args.idesc.desc()));
+  MIOPEN_CHECK(miopenCreateOpConvForward(fusePlanDesc, &convoOp, args.cdesc.desc(), args.wdesc.desc()));
+  MIOPEN_CHECK(miopenCreateOpBiasForward(fusePlanDesc, &biasOp, bdesc.desc()));
+  MIOPEN_CHECK(miopenCreateOpActivationForward(fusePlanDesc, &activOp, miopenActivationRELU));
+
+  // compile fusion plan
+  MIOPEN_CHECK(miopenCompileFusionPlan(args.handle, fusePlanDesc));
+
+  // Set the Args
+  float alpha = static_cast<float>(1);
+  float beta = static_cast<float>(0);
+  float activ_alpha = static_cast<float>(0);
+  float activ_beta = static_cast<float>(0);
+  float activ_gamma = static_cast<float>(0);
+  miopenOperatorArgs_t fusionArgs;
+  MIOPEN_CHECK(miopenCreateOperatorArgs(&fusionArgs));
+  MIOPEN_CHECK(miopenSetOpArgsConvForward(fusionArgs, convoOp, &alpha, &beta, weight.const_data_ptr()));
+  MIOPEN_CHECK(miopenSetOpArgsBiasForward(fusionArgs, biasOp, &alpha, &beta, bias.const_data_ptr()));
+  MIOPEN_CHECK(miopenSetOpArgsActivForward(fusionArgs, activOp, &alpha, &beta, activ_alpha, activ_beta, activ_gamma));
+
+  miopenExecuteFusionPlan(args.handle, fusePlanDesc, args.idesc.desc(), input.const_data_ptr(), args.odesc.desc(), output.data_ptr(), fusionArgs);
+
+  // Cleanup
+  miopenDestroyFusionPlan(fusePlanDesc);
 }
 
 static at::Tensor self_or_new_memory_format(at::Tensor& self, at::MemoryFormat memory_format) {
@@ -1809,107 +1855,171 @@ static at::Tensor self_or_new_memory_format(at::Tensor& self, at::MemoryFormat m
 Tensor miopen_convolution_add_relu(
     const Tensor& input_t,
     const Tensor& weight_t,
-    const Tensor& z_t,
+    const Tensor& z,
     const std::optional<Scalar>& alpha,
-    const std::optional<Tensor>& bias_t,
+    const std::optional<Tensor>& bias,
     IntArrayRef stride,
     IntArrayRef padding,
     IntArrayRef dilation,
     int64_t groups) {
-  auto memory_format = miopen_conv_suggest_memory_format(input_t, weight_t);
-  const Tensor input = input_t.contiguous(memory_format);
-  const Tensor weight = weight_t.contiguous(memory_format);
-  Tensor z = z_t;
-  if (z.suggest_memory_format() != memory_format) {
-    z = z.to(memory_format);
-  }
-  z = z.contiguous(memory_format);
 
-  // FuseFrozenConvAddRelu performs some tensor shape checking
-  Tensor output_t = at::detail::empty_cuda(
-      conv_output_size(
-          input.sizes(), weight.sizes(), padding, stride, dilation),
-      input.options().memory_format(memory_format));
-  if (output_t.numel() == 0) {
-    return output_t;
-  }
+  // MIOpen does not support fusion of add, the alpha2 * z step of the below cuDNN function:
+  // y = act ( alpha1 * conv(x) + alpha2 * z + bias )
+
+  auto memory_format = miopen_conv_suggest_memory_format(input_t, weight_t);
 
   auto& ctx = at::globalContext();
   bool benchmark = ctx.benchmarkCuDNN();
-  auto _alpha = alpha.has_value() ? alpha.value().to<float>() : 1.0;
-  auto _bias = bias_t.has_value()
-      ? bias_t.value()
-      : at::zeros(
-            {output_t.size(1)},
-            optTypeMetaToScalarType(output_t.options().dtype_opt()),
-            output_t.options().layout_opt(),
-            output_t.options().device_opt(),
-            output_t.options().pinned_memory_opt());
 
-  raw_miopen_convolution_add_relu_out(
-      output_t,
+  TensorArg input  { input_t,  "input",  1 },
+            weight { weight_t, "weight", 2 };
+
+  Tensor output_t = at::detail::empty_cuda(
+      conv_output_size(
+        input_t.sizes(), weight_t.sizes(), padding, stride, dilation),
+      input_t.options().memory_format(memory_format));
+  if (output_t.numel() == 0){
+    return output_t;
+  }
+  // Avoid ambiguity of "output" when this is being used as backwards
+  TensorArg output{output_t, "result", 0};
+  miopen_convolution_forward_out(
+      output,
+      "miopen_convolution_add_relu",
       input,
       weight,
-      z,
-      _alpha,
-      _bias,
-      stride,
       padding,
+      stride,
       dilation,
       groups,
       benchmark,
-      true); // deterministic
+      false // deterministic
+  );
 
-  return output_t;
+  auto contig_output_t = self_or_new_memory_format(output_t, memory_format);
+
+  if (!output_t.is_same(contig_output_t)) {
+    contig_output_t.copy_(output_t);
+  }
+
+  auto _alpha = alpha.has_value() ? alpha.value().to<float>() : 1.0;
+  auto _bias = bias.has_value()
+          ? bias.value()
+          : at::zeros(
+                {contig_output_t.size(1)},
+                optTypeMetaToScalarType(contig_output_t.options().dtype_opt()),
+                contig_output_t.options().layout_opt(),
+                contig_output_t.options().device_opt(),
+                contig_output_t.options().pinned_memory_opt());
+
+  at::Tensor alpha_mul_z_add_bias = at::native::reshape_bias(input_t.dim(), _bias).add(z, _alpha);
+  contig_output_t.add_(alpha_mul_z_add_bias);
+  contig_output_t.relu_();
+
+  return contig_output_t;
 }
 
 Tensor miopen_convolution_relu(
     const Tensor& input_t,
     const Tensor& weight_t,
-    const std::optional<Tensor>& bias_t,
+    const std::optional<Tensor>& bias,
     IntArrayRef stride,
     IntArrayRef padding,
     IntArrayRef dilation,
     int64_t groups) {
-  auto memory_format = miopen_conv_suggest_memory_format(input_t, weight_t);
-  const Tensor input = input_t.contiguous(memory_format);
-  const Tensor weight = weight_t.contiguous(memory_format);
-
-  // FuseFrozenConvAddRelu performs some tensor shape checking
-  Tensor output_t = at::detail::empty_cuda(
-      conv_output_size(
-          input.sizes(), weight.sizes(), padding, stride, dilation),
-      input.options().memory_format(memory_format));
-  if (output_t.numel() == 0) {
-    return output_t;
-  }
 
   auto& ctx = at::globalContext();
   bool benchmark = ctx.benchmarkCuDNN();
-  auto _bias = bias_t.has_value()
-      ? bias_t.value()
-      : at::zeros(
-            {output_t.size(1)},
-            optTypeMetaToScalarType(output_t.options().dtype_opt()),
-            output_t.options().layout_opt(),
-            output_t.options().device_opt(),
-            output_t.options().pinned_memory_opt());
 
-  raw_miopen_convolution_add_relu_out(
-      output_t,
-      input,
-      weight,
-      output_t, // use output_t as z to satisfy MIOpen API
-      0, // alpha
-      _bias,
-      stride,
-      padding,
-      dilation,
-      groups,
-      benchmark, // benchmark
-      true); // deterministic
+  // MIOpen currently only supports MemoryFormat::Contiguous and fp32 and 2d
+  if (input_t.suggest_memory_format() == at::MemoryFormat::Contiguous
+          && input_t.scalar_type() == at::kFloat
+          && input_t.ndimension() == 4) {
 
-  return output_t;
+    // FuseFrozenConvAddRelu performs some tensor shape checking
+    Tensor output_t = at::detail::empty_cuda(
+        conv_output_size(
+            input_t.sizes(), weight_t.sizes(), padding, stride, dilation),
+        input_t.options().memory_format(input_t.suggest_memory_format()));
+    if (output_t.numel() == 0) {
+      return output_t;
+    }
+
+    auto _bias = bias.has_value()
+            ? bias.value()
+            : at::zeros(
+                  {output_t.size(1)},
+                  optTypeMetaToScalarType(output_t.options().dtype_opt()),
+                  output_t.options().layout_opt(),
+                  output_t.options().device_opt(),
+                  output_t.options().pinned_memory_opt());
+
+    raw_miopen_convolution_relu_out(
+        output_t,
+        input_t,
+        weight_t,
+        _bias,
+        stride,
+        padding,
+        dilation,
+        groups,
+        benchmark, // benchmark
+        false // deterministic
+    );
+
+    return output_t;
+  }
+  else {
+    // fallback
+
+    auto memory_format = miopen_conv_suggest_memory_format(input_t, weight_t);
+
+    TensorArg input  { input_t,  "input",  1 },
+              weight { weight_t, "weight", 2 };
+
+    Tensor output_t = at::detail::empty_cuda(
+        conv_output_size(
+          input_t.sizes(), weight_t.sizes(), padding, stride, dilation),
+        input->options().memory_format(memory_format));
+    if (output_t.numel() == 0){
+      return output_t;
+    }
+    // Avoid ambiguity of "output" when this is being used as backwards
+    TensorArg output{output_t, "result", 0};
+    miopen_convolution_forward_out(
+        output,
+        "miopen_convolution_relu",
+        input,
+        weight,
+        padding,
+        stride,
+        dilation,
+        groups,
+        benchmark,
+        false // deterministic
+    );
+
+    auto contig_output_t = self_or_new_memory_format(output_t, memory_format);
+
+    if (!output_t.is_same(contig_output_t)) {
+      contig_output_t.copy_(output_t);
+    }
+
+    auto _bias = bias.has_value()
+            ? bias.value()
+            : at::zeros(
+                  {contig_output_t.size(1)},
+                  optTypeMetaToScalarType(contig_output_t.options().dtype_opt()),
+                  contig_output_t.options().layout_opt(),
+                  contig_output_t.options().device_opt(),
+                  contig_output_t.options().pinned_memory_opt());
+
+    at::Tensor reshaped_bias = at::native::reshape_bias(input_t.dim(), _bias);
+    contig_output_t.add_(reshaped_bias);
+    contig_output_t.relu_();
+
+    return contig_output_t;
+  }
 }
 
 REGISTER_CUDA_DISPATCH(miopen_convolution_backward_stub, &miopen_convolution_backward)

aten/src/ATen/native/mps/MPSGraphSonomaOps.h

@@ -0,0 +1,48 @@
+#pragma once
+
+#include <MetalPerformanceShadersGraph/MetalPerformanceShadersGraph.h>
+
+#if !defined(__MAC_14_0) && (!defined(MAC_OS_X_VERSION_14_0) || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_14_0))
+
+typedef NS_ENUM(NSUInteger, MPSGraphFFTScalingMode) {
+  MPSGraphFFTScalingModeNone = 0L,
+  MPSGraphFFTScalingModeSize = 1L,
+  MPSGraphFFTScalingModeUnitary = 2L,
+};
+
+@interface FakeMPSGraphFFTDescriptor : NSObject<NSCopying>
+@property(readwrite, nonatomic) BOOL inverse;
+@property(readwrite, nonatomic) MPSGraphFFTScalingMode scalingMode;
+@property(readwrite, nonatomic) BOOL roundToOddHermitean;
++ (nullable instancetype)descriptor;
+@end
+
+@compatibility_alias MPSGraphFFTDescriptor FakeMPSGraphFFTDescriptor;
+
+@interface MPSGraph (SonomaOps)
+- (MPSGraphTensor* _Nonnull)conjugateWithTensor:(MPSGraphTensor* _Nonnull)tensor name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)realPartOfTensor:(MPSGraphTensor* _Nonnull)tensor name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)fastFourierTransformWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                                      axes:(NSArray<NSNumber*>* _Nonnull)axes
+                                                descriptor:(MPSGraphFFTDescriptor* _Nonnull)descriptor
+                                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)realToHermiteanFFTWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                                    axes:(NSArray<NSNumber*>* _Nonnull)axes
+                                              descriptor:(MPSGraphFFTDescriptor* _Nonnull)descriptor
+                                                    name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)HermiteanToRealFFTWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                                    axes:(NSArray<NSNumber*>* _Nonnull)axes
+                                              descriptor:(MPSGraphFFTDescriptor* _Nonnull)descriptor
+                                                    name:(NSString* _Nullable)name;
+@end
+
+// define BFloat16 enums for MacOS13
+#define MPSDataTypeBFloat16 ((MPSDataType)(MPSDataTypeAlternateEncodingBit | MPSDataTypeFloat16))
+
+// define Metal version
+#define MTLLanguageVersion3_1 ((MTLLanguageVersion)((3 << 16) + 1))
+#endif

aten/src/ATen/native/mps/MPSGraphVenturaOps.h

@@ -0,0 +1,196 @@
+#pragma once
+#include <MetalPerformanceShadersGraph/MetalPerformanceShadersGraph.h>
+
+// TODO: Remove me when moved to MacOS 13
+#if !defined(__MAC_13_2) && (!defined(MAC_OS_X_VERSION_13_2) || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_13_2))
+
+@interface FakeMPSGraphConvolution3DOpDescriptor : NSObject<NSCopying>
+
+@property(readwrite, nonatomic) NSUInteger strideInX;
+@property(readwrite, nonatomic) NSUInteger strideInY;
+@property(readwrite, nonatomic) NSUInteger strideInZ;
+@property(readwrite, nonatomic) NSUInteger dilationRateInX;
+@property(readwrite, nonatomic) NSUInteger dilationRateInY;
+@property(readwrite, nonatomic) NSUInteger dilationRateInZ;
+
+@property(readwrite, nonatomic) NSUInteger paddingLeft;
+@property(readwrite, nonatomic) NSUInteger paddingRight;
+@property(readwrite, nonatomic) NSUInteger paddingTop;
+@property(readwrite, nonatomic) NSUInteger paddingBottom;
+@property(readwrite, nonatomic) NSUInteger paddingFront;
+@property(readwrite, nonatomic) NSUInteger paddingBack;
+
+@property(readwrite, nonatomic) MPSGraphPaddingStyle paddingStyle;
+@property(readwrite, nonatomic) MPSGraphTensorNamedDataLayout dataLayout;
+@property(readwrite, nonatomic) MPSGraphTensorNamedDataLayout weightsLayout;
+
+@property(readwrite, nonatomic) NSUInteger groups;
+
+@end
+
+@compatibility_alias MPSGraphConvolution3DOpDescriptor FakeMPSGraphConvolution3DOpDescriptor;
+
+#endif
+
+@interface MPSGraph (VenturaOps)
+
+#if !defined(__MAC_13_0) && (!defined(MAC_OS_X_VERSION_13_0) || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_13_0))
+
+typedef NS_ENUM(NSUInteger, MPSGraphResizeNearestRoundingMode) {
+  MPSGraphResizeNearestRoundingModeRoundPreferCeil = 0L,
+  MPSGraphResizeNearestRoundingModeRoundPreferFloor = 1L,
+  MPSGraphResizeNearestRoundingModeCeil = 2L,
+  MPSGraphResizeNearestRoundingModeFloor = 3L,
+  MPSGraphResizeNearestRoundingModeRoundToEven = 4L,
+  MPSGraphResizeNearestRoundingModeRoundToOdd = 5L,
+};
+
+// Define complex enums for MacOS 12
+#define MPSDataTypeComplexBit 0x01000000
+#define MPSDataTypeComplexFloat32 ((MPSDataType)(MPSDataTypeFloatBit | MPSDataTypeComplexBit | 64))
+#define MPSDataTypeComplexFloat16 ((MPSDataType)(MPSDataTypeFloatBit | MPSDataTypeComplexBit | 32))
+#endif
+
+- (MPSGraphTensor* _Nonnull)convolution3DWithSourceTensor:(MPSGraphTensor* _Nonnull)source
+                                            weightsTensor:(MPSGraphTensor* _Nonnull)weights
+                                               descriptor:(MPSGraphConvolution3DOpDescriptor* _Nonnull)descriptor
+                                                     name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)
+    convolution3DDataGradientWithIncomingGradientTensor:(MPSGraphTensor* _Nonnull)incomingGradient
+                                          weightsTensor:(MPSGraphTensor* _Nonnull)weights
+                                            outputShape:(MPSShape* _Nonnull)outputShape
+                           forwardConvolutionDescriptor:
+                               (MPSGraphConvolution3DOpDescriptor* _Nonnull)forwardConvolutionDescriptor
+                                                   name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)
+    convolution3DWeightsGradientWithIncomingGradientTensor:(MPSGraphTensor* _Nonnull)incomingGradient
+                                              sourceTensor:(MPSGraphTensor* _Nonnull)source
+                                               outputShape:(MPSShape* _Nonnull)outputShape
+                              forwardConvolutionDescriptor:
+                                  (MPSGraphConvolution3DOpDescriptor* _Nonnull)forwardConvolutionDescriptor
+                                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)cumulativeSumWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                               axis:(NSInteger)axis
+                                               name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                      axis:(NSInteger)axis
+                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                      axis:(NSInteger)axis
+                                descending:(BOOL)descending
+                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                axisTensor:(MPSGraphTensor* _Nonnull)axisTensor
+                                descending:(BOOL)descending
+                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                axisTensor:(MPSGraphTensor* _Nonnull)axisTensor
+                                      name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)argSortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                         axis:(NSInteger)axis
+                                         name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)argSortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                         axis:(NSInteger)axis
+                                   descending:(BOOL)descending
+                                         name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)argSortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                   axisTensor:(MPSGraphTensor* _Nonnull)axisTensor
+                                   descending:(BOOL)descending
+                                         name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)argSortWithTensor:(MPSGraphTensor* _Nonnull)tensor
+                                   axisTensor:(MPSGraphTensor* _Nonnull)axisTensor
+                                         name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)inverseOfTensor:(MPSGraphTensor* _Nonnull)inputTensor name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeNearestWithTensor:(MPSGraphTensor* _Nonnull)imagesTensor
+                                         sizeTensor:(MPSGraphTensor* _Nonnull)size
+                                nearestRoundingMode:(MPSGraphResizeNearestRoundingMode)nearestRoundingMode
+                                       centerResult:(BOOL)centerResult
+                                       alignCorners:(BOOL)alignCorners
+                                             layout:(MPSGraphTensorNamedDataLayout)layout
+                                               name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeNearestWithTensor:(MPSGraphTensor* _Nonnull)imagesTensor
+                                         sizeTensor:(MPSGraphTensor* _Nonnull)size
+                                  scaleOffsetTensor:(MPSGraphTensor* _Nonnull)scaleOffset
+                                nearestRoundingMode:(MPSGraphResizeNearestRoundingMode)nearestRoundingMode
+                                             layout:(MPSGraphTensorNamedDataLayout)layout
+                                               name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeBilinearWithTensor:(MPSGraphTensor* _Nonnull)imagesTensor
+                                          sizeTensor:(MPSGraphTensor* _Nonnull)size
+                                        centerResult:(BOOL)centerResult
+                                        alignCorners:(BOOL)alignCorners
+                                              layout:(MPSGraphTensorNamedDataLayout)layout
+                                                name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeBilinearWithTensor:(MPSGraphTensor* _Nonnull)imagesTensor
+                                          sizeTensor:(MPSGraphTensor* _Nonnull)size
+                                   scaleOffsetTensor:(MPSGraphTensor* _Nonnull)scaleOffset
+                                              layout:(MPSGraphTensorNamedDataLayout)layout
+                                                name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeNearestWithGradientTensor:(MPSGraphTensor* _Nonnull)gradient
+                                                      input:(MPSGraphTensor* _Nonnull)input
+                                        nearestRoundingMode:(MPSGraphResizeNearestRoundingMode)nearestRoundingMode
+                                               centerResult:(BOOL)centerResult
+                                               alignCorners:(BOOL)alignCorners
+                                                     layout:(MPSGraphTensorNamedDataLayout)layout
+                                                       name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeNearestWithGradientTensor:(MPSGraphTensor* _Nonnull)gradient
+                                                      input:(MPSGraphTensor* _Nonnull)input
+                                          scaleOffsetTensor:(MPSGraphTensor* _Nonnull)scaleOffset
+                                        nearestRoundingMode:(MPSGraphResizeNearestRoundingMode)nearestRoundingMode
+                                                     layout:(MPSGraphTensorNamedDataLayout)layout
+                                                       name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeBilinearWithGradientTensor:(MPSGraphTensor* _Nonnull)gradient
+                                                       input:(MPSGraphTensor* _Nonnull)input
+                                                centerResult:(BOOL)centerResult
+                                                alignCorners:(BOOL)alignCorners
+                                                      layout:(MPSGraphTensorNamedDataLayout)layout
+                                                        name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)resizeBilinearWithGradientTensor:(MPSGraphTensor* _Nonnull)gradient
+                                                       input:(MPSGraphTensor* _Nonnull)input
+                                           scaleOffsetTensor:(MPSGraphTensor* _Nonnull)scaleOffset
+                                                      layout:(MPSGraphTensorNamedDataLayout)layout
+                                                        name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sampleGridWithSourceTensor:(MPSGraphTensor* _Nonnull)source
+                                      coordinateTensor:(MPSGraphTensor* _Nonnull)coordinates
+                                                layout:(MPSGraphTensorNamedDataLayout)layout
+                                  normalizeCoordinates:(BOOL)normalizeCoordinates
+                                   relativeCoordinates:(BOOL)relativeCoordinates
+                                          alignCorners:(BOOL)alignCorners
+                                           paddingMode:(MPSGraphPaddingMode)paddingMode
+                                          samplingMode:(MPSGraphResizeMode)samplingMode
+                                         constantValue:(double)constantValue
+                                                  name:(NSString* _Nullable)name;
+
+- (MPSGraphTensor* _Nonnull)sampleGridWithSourceTensor:(MPSGraphTensor* _Nonnull)source
+                                      coordinateTensor:(MPSGraphTensor* _Nonnull)coordinates
+                                                layout:(MPSGraphTensorNamedDataLayout)layout
+                                  normalizeCoordinates:(BOOL)normalizeCoordinates
+                                   relativeCoordinates:(BOOL)relativeCoordinates
+                                          alignCorners:(BOOL)alignCorners
+                                           paddingMode:(MPSGraphPaddingMode)paddingMode
+                                   nearestRoundingMode:(MPSGraphResizeNearestRoundingMode)nearestRoundingMode
+                                         constantValue:(double)constantValue
+                                                  name:(NSString* _Nullable)name;
+- (MPSGraphTensor* _Nonnull)truncateWithTensor:(MPSGraphTensor* _Nonnull)tensor name:(NSString* _Nullable)name;
+
+@end

aten/src/ATen/native/mps/OperationUtils.mm

@@ -9,6 +9,8 @@
 #include <ATen/mps/MPSAllocatorInterface.h>
 #include <ATen/mps/MPSProfiler.h>
 #include <ATen/native/mps/MPSGraphSequoiaOps.h>
+#include <ATen/native/mps/MPSGraphSonomaOps.h>
+#include <ATen/native/mps/MPSGraphVenturaOps.h>
 #include <ATen/native/mps/OperationUtils.h>
 #include <fmt/format.h>
 #include <fmt/ranges.h>
@@ -568,7 +570,7 @@ Placeholder::Placeholder(MPSGraphTensor* mpsGraphTensor,
     MPSShape* mpsStrides = getMPSShape(_tensor.strides());
     check_mps_shape(mpsShape);
 
-    auto storage_numel = src.storage().nbytes() / src.element_size() - src.storage_offset();
+    auto storage_numel = src.storage().nbytes() / src.element_size();
     TORCH_CHECK(storage_numel <= std::numeric_limits<int32_t>::max(),
                 "MPSGaph does not support tensor dims larger than INT_MAX");
     MPSNDArrayDescriptor* srcTensorDesc = [MPSNDArrayDescriptor descriptorWithDataType:dataType

aten/src/ATen/native/mps/kernels/EmbeddingBag.h

@@ -1,25 +0,0 @@
-#pragma once
-#include <c10/metal/common.h>
-
-#ifdef __METAL__
-enum class EmbeddingBagMode { SUM = 0, MEAN, MAX };
-#else
-#include <ATen/native/EmbeddingBag.h>
-using at::native::EmbeddingBagMode;
-#endif
-
-template <typename idx_type_t = uint32_t>
-struct EmbeddingBagParams {
-  ::c10::metal::array<idx_type_t, 2> weight_strides;
-  ::c10::metal::array<idx_type_t, 2> output_strides;
-  ::c10::metal::array<idx_type_t, 2> max_indices_strides;
-
-  idx_type_t per_sample_weights_strides;
-
-  idx_type_t num_indices;
-  idx_type_t num_bags;
-  idx_type_t feature_size;
-
-  EmbeddingBagMode mode;
-  int64_t padding_idx;
-};

aten/src/ATen/native/mps/kernels/EmbeddingBag.metal

@@ -1,212 +0,0 @@
-#include <ATen/native/mps/kernels/EmbeddingBag.h>
-#include <c10/metal/utils.h>
-#include <metal_array>
-#include <metal_stdlib>
-
-using namespace metal;
-using namespace c10::metal;
-
-template <EmbeddingBagMode M, typename T>
-struct ReductionOpInit {
-  inline opmath_t<T> operator()() {
-    return 0;
-  }
-};
-
-template <typename T>
-struct ReductionOpInit<EmbeddingBagMode::MAX, T> {
-  inline opmath_t<T> operator()() {
-    return static_cast<opmath_t<T>>(-INFINITY);
-  }
-};
-
-template <EmbeddingBagMode M, typename T>
-struct ReductionOp {
-  inline opmath_t<T> operator()(
-      T weight_val,
-      opmath_t<T> out_val,
-      uint32_t per_sample_weights_index,
-      constant T* per_sample_weights,
-      uint32_t per_sample_weights_strides);
-};
-
-template <typename T>
-struct ReductionOp<EmbeddingBagMode::SUM, T> {
-  inline opmath_t<T> operator()(
-      T weight_val,
-      opmath_t<T> out_val,
-      uint32_t per_sample_weights_index,
-      constant T* per_sample_weights,
-      uint32_t per_sample_weights_strides) {
-    if (per_sample_weights_strides) {
-      T per_sample_weight = per_sample_weights
-          [per_sample_weights_strides * per_sample_weights_index];
-      return static_cast<opmath_t<T>>(per_sample_weight) *
-          static_cast<opmath_t<T>>(weight_val) +
-          out_val;
-    } else {
-      return static_cast<opmath_t<T>>(weight_val) + out_val;
-    }
-  }
-};
-
-template <typename T>
-struct ReductionOp<EmbeddingBagMode::MEAN, T> {
-  inline opmath_t<T> operator()(
-      T weight_val,
-      opmath_t<T> out_val,
-      uint32_t,
-      constant T*,
-      uint32_t) {
-    return static_cast<opmath_t<T>>(weight_val) + out_val;
-  }
-};
-
-template <typename T>
-struct ReductionOp<EmbeddingBagMode::MAX, T> {
-  inline opmath_t<T> operator()(
-      T weight_val,
-      opmath_t<T> out_val,
-      uint32_t,
-      constant T*,
-      uint32_t) {
-    return max(static_cast<opmath_t<T>>(weight_val), out_val);
-  }
-};
-
-template <EmbeddingBagMode M, typename T>
-struct ReductionOpFinal {
-  inline T operator()(opmath_t<T> val, uint32_t) {
-    return static_cast<T>(val);
-  }
-};
-
-template <typename T>
-struct ReductionOpFinal<EmbeddingBagMode::MEAN, T> {
-  inline T operator()(opmath_t<T> val, uint32_t count) {
-    auto out = val / count;
-    return static_cast<T>((count == 0) ? 0 : out);
-  }
-};
-
-template <typename T>
-struct ReductionOpFinal<EmbeddingBagMode::MAX, T> {
-  inline T operator()(opmath_t<T> val, uint32_t count) {
-    return static_cast<T>((count == 0) ? 0 : val);
-  }
-};
-
-template <EmbeddingBagMode M, typename T, typename I>
-void embedding_bag_impl(
-    constant T* weight,
-    constant I* indices,
-    constant I* offsets,
-    constant T* per_sample_weights,
-    device T* output,
-    device I* offset2bag,
-    device I* bag_size,
-    device I* max_indices,
-    constant EmbeddingBagParams<uint32_t>& params,
-    uint tid) {
-  auto num_indices = params.num_indices;
-  auto num_bags = params.num_bags;
-  auto feature_size = params.feature_size;
-  auto padding_idx = params.padding_idx;
-  auto per_sample_weights_strides = params.per_sample_weights_strides;
-  constant auto& output_strides = params.output_strides;
-  constant auto& weight_strides = params.weight_strides;
-  constant auto& max_indices_strides = params.max_indices_strides;
-
-  auto bag_idx = tid / feature_size;
-  auto feature_idx = tid % feature_size;
-
-  output += bag_idx * output_strides[0] + feature_idx * output_strides[1];
-
-  uint32_t offsets_end = min(bag_idx + 1, num_bags - 1);
-  bool is_last_bag = bag_idx + 1 == num_bags;
-  uint32_t indices_start = static_cast<uint32_t>(offsets[bag_idx]);
-  uint32_t indices_end = is_last_bag * (num_indices) +
-      (!is_last_bag) * (static_cast<uint32_t>(offsets[offsets_end]));
-
-  auto out_val = ReductionOpInit<M, T>()();
-
-  uint32_t bag_size_ = 0;
-
-  for (uint32_t indices_idx = indices_start; indices_idx < indices_end;
-       indices_idx++) {
-    I weight_idx = indices[indices_idx];
-    bool pad = (weight_idx == padding_idx);
-    T weight_val = weight
-        [static_cast<uint32_t>(weight_idx) * weight_strides[0] +
-         feature_idx * weight_strides[1]];
-
-    bag_size_ += static_cast<uint32_t>(!pad);
-
-    auto tmp_val = ReductionOp<M, T>()(
-        weight_val,
-        out_val,
-        indices_idx,
-        per_sample_weights,
-        per_sample_weights_strides);
-
-    out_val = pad ? out_val : tmp_val;
-  }
-
-  *output = ReductionOpFinal<M, T>()(out_val, bag_size_);
-}
-
-#define DISPATCH_IMPL(MODE)        \
-  return embedding_bag_impl<MODE>( \
-      weight,                      \
-      indices,                     \
-      offsets,                     \
-      per_sample_weights,          \
-      output,                      \
-      offset2bag,                  \
-      bag_size,                    \
-      max_indices,                 \
-      params,                      \
-      tid)
-
-template <typename T, typename I>
-kernel void embedding_bag(
-    constant T* weight [[buffer(0)]],
-    constant I* indices [[buffer(1)]],
-    constant I* offsets [[buffer(2)]],
-    constant T* per_sample_weights [[buffer(3)]],
-    device T* output [[buffer(4)]],
-    device I* offset2bag [[buffer(5)]],
-    device I* bag_size [[buffer(6)]],
-    device I* max_indices [[buffer(7)]],
-    constant EmbeddingBagParams<uint32_t>& params [[buffer(8)]],
-    uint tid [[thread_position_in_grid]]) {
-  switch (params.mode) {
-    case EmbeddingBagMode::SUM:
-      DISPATCH_IMPL(EmbeddingBagMode::SUM);
-    case EmbeddingBagMode::MEAN:
-      DISPATCH_IMPL(EmbeddingBagMode::MEAN);
-    case EmbeddingBagMode::MAX:
-      DISPATCH_IMPL(EmbeddingBagMode::MAX);
-  }
-}
-
-#define REGISTER_EMBEDDING_BAG_OP(T, I)                             \
-  template [[host_name("embedding_bag_" #T "_" #I)]]                \
-  kernel void embedding_bag<T, I>(                                  \
-      constant T * weight [[buffer(0)]],                            \
-      constant I * indices [[buffer(1)]],                           \
-      constant I * offsets [[buffer(2)]],                           \
-      constant T * per_sample_weights [[buffer(3)]],                \
-      device T * output [[buffer(4)]],                              \
-      device I * offset2bag [[buffer(5)]],                          \
-      device I * bag_size [[buffer(6)]],                            \
-      device I * max_indices [[buffer(7)]],                         \
-      constant EmbeddingBagParams<uint32_t> & params [[buffer(8)]], \
-      uint tid [[thread_position_in_grid]]);
-
-REGISTER_EMBEDDING_BAG_OP(float, int);
-REGISTER_EMBEDDING_BAG_OP(float, long);
-REGISTER_EMBEDDING_BAG_OP(half, int);
-REGISTER_EMBEDDING_BAG_OP(half, long);
-REGISTER_EMBEDDING_BAG_OP(bfloat, int);
-REGISTER_EMBEDDING_BAG_OP(bfloat, long);

aten/src/ATen/native/mps/operations/BinaryKernel.mm

@@ -8,6 +8,8 @@
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/mps/OperationUtils.h>
 #include <ATen/native/mps/operations/BinaryKernel.h>
+// For MTLLanguageVersion_3_1
+#include <ATen/native/mps/MPSGraphSonomaOps.h>
 #include <fmt/format.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS

aten/src/ATen/native/mps/operations/Convolution.mm

@@ -1,12 +1,23 @@
 //  Copyright © 2022 Apple Inc.
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/native/ConvUtils.h>
+#include <ATen/native/mps/MPSGraphVenturaOps.h>
 #include <ATen/native/mps/OperationUtils.h>
 #include <ATen/ops/_mps_convolution_native.h>
 #include <ATen/ops/_mps_convolution_transpose_native.h>
 #include <ATen/ops/mps_convolution_backward_native.h>
 #include <ATen/ops/mps_convolution_transpose_backward_native.h>
-#include <fmt/format.h>
+
+#if !defined(__MAC_13_2) && (!defined(MAC_OS_X_VERSION_13_2) || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_13_2))
+
+@implementation FakeMPSGraphConvolution3DOpDescriptor
+- (nonnull id)copyWithZone:(nullable NSZone*)zone {
+  return self;
+}
+
+@end
+
+#endif
 
 namespace at::native {
 
@@ -39,9 +50,11 @@ static void fill_conv3d_desc(MPSGraphConvolution3DOpDescriptor* descriptor_,
   descriptor_.paddingFront = paddingDepth;
   descriptor_.paddingBack = paddingDepth;
 
-  descriptor_.dataLayout = MPSGraphTensorNamedDataLayoutNCDHW;
+  // PyTorch always uses NCDHW memory layout for 3D tensors
+  descriptor_.dataLayout = (MPSGraphTensorNamedDataLayout)7L; // MPSGraphTensorNamedDataLayoutNCDHW;
 
-  descriptor_.weightsLayout = MPSGraphTensorNamedDataLayoutOIDHW;
+  // PyTorch always uses OIDHW memory layout for 3D weights
+  descriptor_.weightsLayout = (MPSGraphTensorNamedDataLayout)9L; // MPSGraphTensorNamedDataLayoutOIDHW;
 
   descriptor_.groups = groups; // not yet tested in Xcode/C++
 }
@@ -173,6 +186,18 @@ static Tensor _mps_convolution_impl(const Tensor& input_t_,
     if (bias_defined)
       bias_shape = bias_opt.value().sizes();
 
+    std::string mem_format_key;
+    switch (memory_format) {
+      case at::MemoryFormat::Contiguous:
+        mem_format_key = "Contiguous";
+        break;
+      case at::MemoryFormat::ChannelsLast:
+        mem_format_key = "ChannelsLast";
+        break;
+      default:
+        assert(0 && "Check should have been done earlier\n");
+    }
+
     std::string bias_shape_key;
     if (bias_defined) {
       bias_shape_key = std::to_string(bias_shape[0]);
@@ -180,16 +205,20 @@ static Tensor _mps_convolution_impl(const Tensor& input_t_,
       bias_shape_key = "nobias";
     }
 
-    std::string key = fmt::format("mps_{}convolution:{}:{}:{}:{}:{}:{}:{}:{}",
-                                  is3DConv ? "3d_" : "",
-                                  getArrayRefString(stride),
-                                  getArrayRefString(dilation),
-                                  getArrayRefString(padding),
-                                  groups,
-                                  is_channels_last,
-                                  mps::getTensorsStringKey({input_t, weight_t}),
-                                  bias_defined,
-                                  bias_shape_key);
+    std::string key;
+    if (is3DConv) {
+      key = "mps_3d_convolution:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          std::to_string(stride[2]) + ":" + std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" +
+          std::to_string(dilation[2]) + ":" + std::to_string(padding[0]) + ":" + std::to_string(padding[1]) + ":" +
+          std::to_string(padding[2]) + ":" + std::to_string(groups) + ":" + mem_format_key +
+          mps::getTensorsStringKey({input_t, weight_t}) + ":" + std::to_string(bias_defined) + ":" + bias_shape_key;
+
+    } else {
+      key = "mps_convolution:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" + std::to_string(padding[0]) + ":" +
+          std::to_string(padding[1]) + ":" + std::to_string(groups) + ":" + mem_format_key +
+          mps::getTensorsStringKey({input_t, weight_t}) + ":" + std::to_string(bias_defined) + ":" + bias_shape_key;
+    }
 
     MPSShape* inputShape = mps::getMPSShape(input_t, memory_format);
     MPSShape* outputShape = mps::getMPSShape(output_t, memory_format);
@@ -371,15 +400,33 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,
   @autoreleasepool {
     MPSStream* stream = getCurrentMPSStream();
 
+    std::string mem_format_key;
+    switch (memory_format) {
+      case at::MemoryFormat::Contiguous:
+        mem_format_key = "Contiguous";
+        break;
+      case at::MemoryFormat::ChannelsLast:
+        mem_format_key = "ChannelsLast";
+        break;
+      default:
+        assert(0 && "Check should have been done earlier\n");
+    }
+
     MPSShape* mps_input_shape = getMPSShape(input_size);
-    std::string key = fmt::format("mps_{}_convolution_backward_input:{}:{}:{}:{}:{}:{}",
-                                  is3DConv ? "3d_" : "",
-                                  getArrayRefString(stride),
-                                  getArrayRefString(dilation),
-                                  getArrayRefString(padding),
-                                  groups,
-                                  is_channels_last,
-                                  getTensorsStringKey({grad_output_t, weight_t}));
+    std::string key;
+    if (is3DConv) {
+      key = "mps_3d_convolution_backward_input:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          ":" + std::to_string(stride[2]) + std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" +
+          std::to_string(dilation[2]) + ":" + std::to_string(padding[0]) + ":" + std::to_string(padding[1]) + ":" +
+          std::to_string(padding[2]) + ":" + std::to_string(groups) + ":" + mem_format_key +
+          getTensorsStringKey({grad_output_t, weight_t});
+
+    } else {
+      key = "mps_convolution_backward_input:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" + std::to_string(padding[0]) + ":" +
+          std::to_string(padding[1]) + ":" + std::to_string(groups) + ":" + mem_format_key +
+          getTensorsStringKey({grad_output_t, weight_t});
+    }
     auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
       auto gradOutputTensor = mpsGraphRankedPlaceHolder(mpsGraph, grad_output_t);
       auto weightTensor = mpsGraphRankedPlaceHolder(mpsGraph, weight_t);
@@ -504,13 +551,19 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
     MPSStream* stream = getCurrentMPSStream();
 
     MPSShape* mps_weight_shape = getMPSShape(weight_size);
-    std::string key = fmt::format("mps_{}convolution_backward_weights:{}:{}:{}:{}:{}",
-                                  is3DConv ? "3d_" : "",
-                                  getArrayRefString(stride),
-                                  getArrayRefString(dilation),
-                                  getArrayRefString(padding),
-                                  groups,
-                                  getTensorsStringKey({grad_output_t, input_t, grad_weight_t}));
+    std::string key;
+    if (is3DConv) {
+      key = "mps_3d_convolution_backward_weights:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          std::to_string(stride[2]) + ":" + std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" +
+          std::to_string(dilation[2]) + ":" + std::to_string(padding[0]) + ":" + std::to_string(padding[1]) + ":" +
+          std::to_string(padding[2]) + ":" + std::to_string(groups) + ":" +
+          getTensorsStringKey({grad_output_t, input_t, grad_weight_t});
+    } else {
+      key = "mps_convolution_backward_weights:" + std::to_string(stride[0]) + ":" + std::to_string(stride[1]) + ":" +
+          std::to_string(dilation[0]) + ":" + std::to_string(dilation[1]) + ":" + std::to_string(padding[0]) + ":" +
+          std::to_string(padding[1]) + ":" + std::to_string(groups) + ":" +
+          getTensorsStringKey({grad_output_t, input_t, grad_weight_t});
+    }
     auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
       MPSShape* inputShape = getMPSShape(input_t);
       bool isDepthwiseConv =

aten/src/ATen/native/mps/operations/Copy.mm

@@ -2,6 +2,7 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/mps/MPSProfiler.h>
 #include <ATen/native/mps/Copy.h>
+#include <ATen/native/mps/MPSGraphSonomaOps.h>
 #include <ATen/native/mps/OperationUtils.h>
 #include <ATen/ops/_copy_from_and_resize_native.h>
 #include <ATen/ops/_copy_from_native.h>

aten/src/ATen/native/mps/operations/Distributions.mm

@@ -5,6 +5,8 @@
 #include <ATen/native/DistributionTemplates.h>
 #include <ATen/native/Distributions.h>
 #include <ATen/native/TensorFactories.h>
+#include <ATen/native/mps/MPSGraphSonomaOps.h>
+#include <ATen/native/mps/MPSGraphVenturaOps.h>
 #include <ATen/native/mps/OperationUtils.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS